Engineering portfolio · chief-engineer edition · director edition

Nir Herscovici

Head of Manufacturing Technology and Principal Engineer (R&D), Precision Castparts Corp. (Berkshire Hathaway), 2020–2026. Architect of the perception and control platform that runs 24/7 at eight or more aerospace-metals plants, and the engineer who wrote production code for the systems on it.

With the company 2014–2026 · M.S. and B.S. Mechanical Engineering, B.S. Computer Science, University of Nevada, Las Vegas · Co-inventor on four patent filings, one granted

  • 8+plants on one reference kit: 200+ sensors, 30 GPU servers, zero production downtime
  • 15+systems in 24/7 production, each with an assurance case before fielding
  • 4patent filings, one granted, all in production
  • 30%+less scrap and defects across the plants running these systems
  • 22person R&D organization built and led for six years
RolesHead of Manufacturing Technology and Principal Engineer, R&DSenior Manufacturing Process Technology EngineerResearch and Data Analytics Engineer, R&DManufacturing Process Technology Engineer, R&D2014201620182020202220242026MilestonesVision program starts,Univ. of BirminghamFirst in-houseend-of-melt visionNeural-networkdevelopment beginsHead of a 22-personR&D organizationProvisional patent,power cutbackLMPC Best PaperAward$500M greenfieldplant breaks groundLMPC Most InnovativeWork AwardU.S. patent granted;two more publishedWeld patent filed;conference paper
  1. 2014Manufacturing Process Technology Engineer, R&D. Vision program started with the University of Birmingham
  2. 2015Research and Data Analytics Engineer, R&D
  3. 2016First in-house end-of-melt vision system
  4. 2017Senior Manufacturing Process Technology Engineer
  5. 2018Neural-network development begins
  6. 2020Head of Manufacturing Technology and Principal Engineer, R&D: a 22-person R&D organization
  7. 2021Provisional patent, power cutback
  8. 2022LMPC Best Paper Award
  9. 2023$500M greenfield plant breaks ground
  10. 2024LMPC Most Innovative Work Award
  11. 2025U.S. patent granted; two applications published
  12. 2026Weld-rating patent filed; conference paper
Career progression at Precision Castparts, 2014–2026.

Part 1 · Scope

Leadership and scope

Herscovici ran manufacturing technology for melting, forging and inspection across the company's titanium plants: a 22-person R&D organization, a seven-figure budget, the company's de facto head of machine learning, and the technology standard that every site builds to. He presented roadmaps directly to the COO, won CEO approval to expand the organization, and was the company's technical face to nine jet-engine OEMs through a material-defect investigation.

Scope

Perception and control for melting, forging and inspection at 8+ plants in the U.S. and UK. 15+ systems in 24/7 production, 200+ sensors on 30 GPU servers, zero production downtime through every rollout.

People

10 engineers, 5 graduate researchers, 7 technicians. Hired four engineers directly, recruited three internally, advised all five researchers. Every engineer earned a significant promotion; one technician is now a superintendent; 85% of 15 interns became full-time engineers.

Money

Seven-figure R&D budget. Seven-figure capital requests approved on business cases with payback targets typically under 12 months. $2M+ of new OEM and cross-division funded work. Owner of the $5M+ technology package at a $500M greenfield plant.

Stakeholders

COO (roadmaps and schedules), CEO (organization expansion), OEM executives and specialists (weekly during the investigation), three engineering contractors and the furnace OEM (greenfield build), plant IT (policy standards), a DARPA consortium with four OEMs.

Title note: PCC has no chief-engineer title. "De facto chief engineer" describes the role he held at the greenfield plant, not a title.

Part 2 · Architecture and the technology roadmap

Build once, then scale across plants and repurpose across processes

Herscovici set the multi-year technology direction for machine learning, vision, sensing, inspection and automation across the company's plants and processes. The roadmap matured from isolated AI projects into a manufacturing-systems platform, and from one plant's tools into a catalog of capabilities: each one built and hardened on one process, then copied to other plants and repurposed for other processes and divisions. The model is a standard core with a local last mile: the hardware, software and data architecture are shared; mounting, thresholds, operator displays and process logic are site-specific.

SensingDetectionMeasurementDataintegrationControlsStorageand retrievalOperatorworkflowRetrainingand supportDeploymentgovernancewhere an isolated project stopswhat makes it a production system that can be copied and repurposedFrom isolated AI projects to a manufacturing-systems platformstandard core, local last mile: mounting, thresholds, displays and process logic are site-specific; everything else is shared
FIG. 1The roadmap as a chain. Isolated projects stop at measurement; the six links after it are what turn a tool into a production system that can be copied and repurposed.

The same failure modes repeated across facilities: manual supervision, fragile data integration, underestimated hardware complexity, fragmented ownership and reinvention. Each became a standardization priority: preferred hardware stacks, reusable software modules, a signal and storage architecture, documentation and acceptance-test packages, IT and security templates, and named ownership for updates and retraining, so no deployment is engineered from scratch.

Build once: what the kit has to meet

  • Survive the furnace. Radiant heat, arc light, metal particulate, vibration, vacuum interfaces, and electrical noise over 100 to 300 m of cable from furnace to control room.
  • Never remove the operator's live view. No single failure in analytics or storage may take it away.
  • Act on the process without interrupting it. New systems connect while furnaces run and need no manual tuning; a closed loop works to a budget of 1 to 5 seconds from frame capture to PLC action.
  • Fail safe. Where a result gates a step, a missing result stops the step; the operator is alarmed before any automatic action.
  • Reuse at any site. One hardware, network, compute and integration standard, under IT policy agreed in advance.
  • Maintainable by plant staff. No licensed black boxes, no outside specialists for routine changes.

The reference kit

SensingTransportComputePlant systemsPeopleCameras150, ruggedized housings,vacuum-rated viewportsThermal imagers20Laser measurement35 systemsThe processpower, welder release,arc-steering coil currentThree independentpaths per camera1 SDI over fiberCWDM, EMI-immune2 Ethernetto recorders3 HDMIto operator displaysPerception modelsdetection, segmentation,tracking, measurement,decision logicon 30 rack-mountedGPU servers, 24/7Video historian24/7 recording; the sourceof training dataAssuranceacceptance gates, calibratedthresholds, drift monitorsPlant historianratings, measurements,process dataMESheat and billet recordsPLCs over OPC-UAinterlocks, power cutback,coil currentEngineersevent and record reviewOEM customersevidence from reviewedeventsOperatorslive video, alarms,decision support123 HDMI: live video straight to the operator displaysconfirmed events retrain the modelsclosed loop: alarm → decision support → automatic action if no one respondsGovernanceone hardware and vendor standard across 8+ plants · factory acceptance testing · rollout playbook and maturity levelsmaintenance and troubleshooting procedures · assurance gates before fielding · zero production downtime
FIG. 2Sensor to process. Red: the closed loop, in which a model's result reaches the PLC and acts on process power, welder release or coil current, with the operator alarmed first. The governance band is what makes the same design run at every plant.

Sensing: 150 cameras, 20 thermal imagers, 35 laser measurement systems, five to eight cameras per furnace at the greenfield plant; 25+ units in shielded, water-cooled, purge-air housings behind vacuum-rated viewports. Transport: three independent outputs per camera, SDI over fiber (CWDM, several 3G and 12G-SDI feeds per strand, a 40×40 video router) to analytics, Ethernet to recorders, HDMI to the operator displays. Compute: 30 rack-mounted GPU servers running detection, segmentation, tracking, measurement and decision logic, with a video historian that records around the clock to cloud storage and became the training-data source for every later model. Integration: PLCs over OPC-UA, the MES and the plant historian; two interface points per system. People: operator displays and alarms, engineering review of recorded events.

As-built vision signal layout drawing from camera through two junction boxes to analytics, storage and live video.
FIG. 3The as-built signal layout (version 7, 2023). 1 One camera; its SDI, HDMI and Ethernet outputs go into the furnace-side junction box. 2 One fiber, 100 to 300 m, is the only thing that leaves the furnace area. 3 In the control-room cabinet the signals split again: analytics over SDI, storage and camera control over Ethernet. 4 Live video reaches the operator's monitor over HDMI on its own path.
Photograph of a camera in a shielded housing on a mast at a furnace, with a light, conduit and a junction box below.
FIG. 4A camera as installed at a furnace. 1 The camera in its shielded housing. 2 The light on the same mast. 3 Conduit carrying the three outputs down to 4 the junction box, where they become one fiber.
Photograph looking into a camera housing: a compact camera on a bracket, with its cable, behind a round viewport.
FIG. 5Behind the viewport. 1 The camera body, a compact industrial camera. 2 Its mounting plate, which sets the optical axis to the viewport. 3 The service cable to the junction box. 4 The rim of the viewport the camera looks through. The housing takes the heat, purge and shielding so the camera does not have to.
Photograph of an open junction box rack with converters, terminations and a 120 volt label.
FIG. 6The furnace-side junction box. 1 SDI-to-fiber converters, one per camera output. 2 Terminations and cable management; the single fiber leaves from here. 3 120 V service, the only copper that stays on the furnace side.
Photograph of an open control-room cabinet with a fiber patch panel, an SDI distribution unit, a recorder, a switch and breakers.
FIG. 7The control-room cabinet for one furnace. 1 Fiber patch panel: every camera on this furnace arrives here. 2 SDI distribution to the analytics servers and router. 3 Recorder and network switch. 4 Protected power.
CAD view of the top of an electron-beam furnace with camera stations, thermal imagers and a level sensor arranged around the chamber.
FIG. 8Sensor placement engineered in CAD before installation, around an electron-beam furnace. 1 One of seven camera stations (ST1 to ST7), each with its viewport and clearances. 2 One of two thermal imagers (TM2). 3 The level sensor (LC1).

Design rules

  • Standardize the hardware so scaling is routine. A new site reuses the stack instead of designing from scratch.
  • Three paths per camera. Redundant signals reduce failure points; the operator's view survives any analytics fault.
  • Two interface points per system. Minimal coupling to the process controls; changes while the furnace runs.
  • Operator in the loop, automatic fallback. Alarm, decision support, then action only if no one responds. The PLC, not the model, holds every interlock.
  • Measure against a second reference. Camera against laser on the same part, detections against injected events, ratings against an engineer.
  • Networks trace, geometry decides. Where a number gates a process, a neural network finds the part and deterministic geometry produces the number, so the result generalizes and can be audited.

Verification and rollout

  • Factory acceptance testing before installation: every camera and fiber link powered and imaged end to end; router, recorders and servers configured and load-tested; the analytics models run on recorded video on the production servers; PLC and MES interfaces exercised with the furnace OEM. Electrical schematics, cable labeling and control-room display layouts owned by the function.
  • Model updates are validated offline on recorded video, then deployed in a maintenance window; video is stored in the cloud, so the record is not retention-limited.
  • Assurance gates before any model is fielded: FMEA and hazard analysis with documented safety cases, acceptance criteria, calibrated confidence thresholds, runtime input-quality and drift monitors.
  • Detection validated against injected ground-truth events, 95%+ detected, before fielding.
  • Greenfield plant: seven furnaces instrumented from site inception, milestones met. Second plant: six months with zero reported failures. Across the platform: zero production downtime through every rollout.

Scale: copy to the next plant

A new site does not get a new design. It gets the reference kit, the same parts, drawings and procedures, and moves through the same four maturity stages. That is how one system on one furnace in 2016 became a company standard at eight or more plants by 2026, with a greenfield plant instrumented from inception. Read-across is a business lever as much as an engineering one: less repeated engineering cost, shorter deployment time, and site teams that trust what arrives because it already runs elsewhere.

One designMany instancesReference designsensingtransportcomputeintegrationgovernanceinstantiatesite 1site 2site 3site 4site nsolid: the shared core, identical everywhere · hatched: the local last milegovernance: updates flow down, lessons flow upGrowth20161 unit20205 units20237 units, greenfield20268+ plants
FIG. 9Scaling as instantiation. One reference design; each site is a copy with a small site-specific slot; governance keeps the copies in step. Red: how the count grew, from one unit in 2016 to eight or more plants in 2026.

Repurpose: the same capability on a different process

The capabilities are not tied to the process they were built on. Rare-event detection trained on one process became defect and side-arcing recognition for a castings division; the measurement stack became billet straightness and cut guidance at a forge press and CT defect detection for castings; process analytics moved to an electron-beam furnace. Each move reused validated architecture and software patterns, which is what makes the portfolio a corporate asset rather than a set of plant projects.

CapabilityHow far it traveled from where it was builtSame process,more plantsDifferentprocessDifferentdivisionDifferent sensingmodalityRare-event detection and trackingvideo models, tracking, evidence pipelineClosed-loop control from visionalarm, decision support, automatic actionGated measurementa number the controller can act onMeasurement and sensor fusiontwo instruments, one recordElectromagnetic actuation and controlcoils, drives, controller, verification by cameraProcess analytics and safetyquality scoring, hazard zonesdoneon the roadmap, same kitnot applicableblank: not yetA capability that has crossed a column boundary once has a template for crossing it again.
FIG. 10How far each capability traveled from the process it was built on: to more plants, to a different process, to a different division, to a different sensing modality. Filled: done. Rings: on the roadmap with the same kit.

Part 3 · Case studies

Case studies: the engineering in depth

Six systems, written for an engineer: the decisions and what they cost, the sketches, the governing mathematics, the measured impact and the limits. Herscovici wrote production code for each of them, in Python with PyTorch, OpenCV and scikit-learn, with Databricks and Spark for the investigation pipeline and OPC-UA for the controller integration, alongside the engineers who now maintain them.

Data

Rare-event strategy with confirmed events fed back into training. Segment Anything fine-tuned on 1,000 to 10,000 labeled images per task and used to mass-label the datasets that train other models.

Evaluation

Injected ground-truth events (95%+ detected); per-stage false-positive counts against known events; ratings benchmarked against an engineer; camera measurements benchmarked against a laser instrument; a domain-shift monitor that rates each video against the training set.

Deployment

30 rack-mounted GPU servers and edge devices; models pruned 50%+ in size with better accuracy on under-represented data; ONNX and TensorRT; Databricks, Spark and MLflow for the large reviews.

Operations

Assurance gates before fielding; runtime input-quality and drift monitors; a video historian as the training source; on-call and incident response for systems in service.

Expand all · The code is under NDA; the patents and the paper are the public record.

Machine perception and rare-event detection

01 Rare-event detection and tracking in production video (image-suitability method: U.S. application pending)Resolved a first-of-its-kind jet-engine material defect for nine OEMs, measured by 95% fewer false positives at 95%+ recall, 100× less manual review and an ~80% lower projected review cost, by designing a three-stage detector with multi-target tracking and a domain-shift monitor, and running the review on a weekly schedule as customer-facing technical lead.

His part: customer-facing technical lead; designed and coded the detection and tracking system; directed 50+ people. Scale: 1,132 runs, 7 customer groups, 2,200 videos, 4 TB. Customers: Pratt & Whitney, GE Aerospace, Rolls-Royce, Lockheed Martin and five other OEMs.

Problem and constraints

Seven consecutive video frames of the glowing annulus between electrode and crucible with a small particle crossing it.
FIG. 11Seven consecutive frames from a furnace camera, zoomed progressively. 1 A particle a few pixels across enters the gap between electrode and crucible and 2 crosses it in a handful of frames. Its path over the sequence is what tells electrode-origin from crucible-origin debris.

Debris that falls into the melt can leave a defect in the ingot. It enters through the narrow gap between the electrode and the crucible, stays visible for a few frames, and can appear at any point in a melt that runs for hours. Only 73 confirmed events existed for calibration; lighting and camera quality varied across furnaces and years; and every flagged event had to hold up as evidence for customers.

Decisions and trade-offs

  • Three small stages instead of one large model. About eight detector variants were benchmarked per melt stage. A tiled nano detector with a high-resolution P2 head, a random forest on a 2,242-feature crop and a temporal filter kept recall on few-pixel particles at a fraction of the compute of a single large model, at the cost of three models to maintain.
  • Rate the data, not only the model. A domain-shift monitor scores each video against the training distribution. It showed most field video was darker than the training images, which pointed the fix at the camera image rather than at more training.
  • An evidence standard over speed. Every flagged event carries a timestamp, a saved frame and an engineering review before it reaches a customer. It slowed the pipeline, and it is why the results held with nine OEMs.

Solution

three stages remove false positives before reviewFurnace videoproduction runs,with telemetryFrames, tilesfull resolutionfor small targetsStage 1 · DetectorYOLO26n with ahigh-resolution P2 headStage 2 · Classifierrandom forest on 2,242texture and shape featuresStage 3 · Filterchecks detectionsagainst earlier frameseventsTrackingtrajectory gives originRisk scoresize against the gapEngineering reviewframe + timestampTraining setconfirmed eventsretrain
FIG. 12The pipeline. Red bracket: the three stages that remove false positives before anything reaches an engineer. Confirmed events flow back into training.
Text panel of an event record: detection time 4.872 s, detector confidence 0.589, localized-change score 0.902, mean and 95th-percentile difference, changed-pixel fractions and inside-to-ring ratio.
FIG. 13One event as the pipeline records it: detection time, detector confidence, the localized-change score and the change statistics behind it. Every flagged event carries this record, a saved frame and an engineering review before it reaches a customer.

Multi-target tracking follows each particle frame to frame; its trajectory tells debris shed by the electrode from debris released by the crucible wall. Because camera zoom differs between furnaces, size is measured against the gap, whose width is known. Melt-generated debris carries no weight; other particles are weighted by size, doubled when the trajectory starts at the electrode weld:

The domain-shift monitor samples 150 frames per video, finds the melt region M and computes its size A, sharpness S and brightness B; each video is represented by the 25th percentile of its frame metrics, scaled by the training interquartile range and averaged into one distance:

Bar chart of training-set coverage relative to the application distribution for sharpness, brightness and annulus area; the brightness above-p90 bar is 72%.
FIG. 14Why the field video rated low. 1 72% of training images were brighter than the 90th percentile of field video: the training set was lit better than the plant. 2 Sharpness and melt-region area were covered. The finding pointed the fix at the camera image, not at more training.
Bar chart of video-level rating distribution: 84 at rating 1, 2,271 at rating 2, 1,382 at rating 3.
FIG. 15The domain-shift monitor over 3,737 field videos from 97 production runs. 1 84 videos rated 1, outside the training distribution. 2 2,271 rated 2 and 1,382 rated 3; none rated 4 or 5.

Video was fused with timestamp-aligned controller telemetry and processed on Databricks. The models were trained on 30,000+ hours of production video. A weekly schedule tracked all 1,132 runs through export, upload, processing, engineering review and group review against plan.

Impact

Manual review fell 100-fold, with 95% fewer false positives at 95%+ recall; the three-stage detector replaced manual review of raw video. The investigation resolved the defect. The pipeline now serves as an investigative tool for past melts and has reviewed more than 1,000 hours of video.

Limits

Calibration rested on 73 confirmed events. Field video rated 2 or 3 on the domain-shift scale sits outside the center of the training distribution; the monitor flags that, it does not fix it. Recall is stated on confirmed events, not on events no reviewer saw.

Vision systems for dimensional metrology and inspection

02 A measurement that gates a process step: weld rating (provisional patent filed)Replaced a visual weld judgment with a reproducible gate on the melt, measured by a 1-to-5 rating in 5–10 seconds, ~85% agreement with an engineer mid-development and a PLC interlock on every failed weld, by building two segmentation networks, a radial-coverage geometry and a fail-closed PLC and MES integration.

His part: led and coded the system; co-inventor on the provisional patent application; co-author of the 2026 PCC Technical Conference paper. Period: 2025–2026.

Problem and constraints

Before an electrode is melted, the weld joining it to its stub is rated, because weld material can break away and fall into the melt; a particle from the weld carries twice the risk weight of any other in case study 01. Operators rated the weld by eye, so it varied between shifts. The scene is a shop floor with people, equipment and lighting from dark to bright; the result is needed within seconds; a missing answer must stop the process; and every rating must be traceable to heat, electrode and operator.

Decisions and trade-offs

  • Segmentation and geometry instead of a classifier. A first version rated welds with a bounding-box network and could only recognize conditions it had been trained on. The second traces the weld and the electrode and rates by coverage geometry, which costs a second network and generalizes: a deliberately bad weld outside the training set was rated correctly.
  • Two cameras, not one. The top sector of the electrode is poorly visible from one viewpoint, so each welder has two cameras about 180° apart, each rating the sector it sees well.
  • The PLC holds the interlock and the system fails closed. No electrode edge found means no rating and a locked welder. A software-only gate that passes on uncertainty was rejected.

Solution

StationSystemCamera 1Camera 2LightLightStubWeld pool:what is ratedElectrodeWelderchambercameras about 180° apartGPU workstationweld CNN + electrode CNNgeometric rating, 5–10 sOperator interfaceoperator rating, result,both camera viewsRecorder24/7 videoWelder PLCheat and electrode IDsinterlockMES and historianrating record per heatHD video (SDI)IDs in,rating outInterlock: locks the welder,blocks the melt
FIG. 16The station. Red, left: the weld pool at the base of the stub, which is what is rated. Red arrow: the PLC interlock that locks the welder and blocks the melt.
CAD render of a stub-welder structure with camera pairs mounted on the frame around the weld station and a junction box at the base.
FIG. 17The welder vision installation as designed: camera pairs on the welder frame around the weld station, lights, and the junction box at the base.

One CNN traces the weld contour and a lighter CNN finds the electrode boundary. A best-fit ellipse to the electrode mask gives a center; rays cast from it through 360° record the furthest electrode point re(θ) and the furthest weld point rw(θ) on each ray. Their ratio is the coverage, and the furthest protrusion sets the rating:

Photograph of an electrode on a stub welder with the system overlay: the weld traced in red, concentric rating bands, and a marked protrusion labeled rating 3.
FIG. 18The system's own overlay on a live weld. 1 The weld contour traced by the segmentation network. 2 The rating bands 2 to 5 drawn on the fitted electrode ellipse. 3 The furthest protrusion, where the weld crosses the band that sets the rating (here 3). 4 The per-ray readout: gap, r_w, r_e and coverage.
Image geometry2345θCr_e(θ): electrodeedge on this rayr_w(θ): furthest weldpoint on this rayStub (hatched)Weld contour(segmentation CNN)Best-fit ellipse Eof the electrode faceTop sector ignored:rated by the second cameraDashed ellipses: rating bands 2 to 5 at coverage 0.60, 0.75, 0.90 and 0.97Coverage around the electrodeignored sector20.6030.7540.9050.97rating0°90°180°270°360°angle θcov(θ)peak 0.93: rating 4
FIG. 19The geometry of a rating. Red contour and point: the weld protrusion and its furthest reach along one ray. Red circle, right: the same protrusion as the peak of the coverage curve, crossing the 0.90 band for rating 4.
Line plot of weld coverage versus angle from 0 to 360 degrees with dashed threshold lines for ratings 2 to 5; the curve peaks near 0.91 around 140 degrees.
FIG. 20Coverage against ray angle for one weld, from the system. 1 The peak, 0.91, crosses the 0.90 band and sets the rating at 4. 2 The 0.60 band for rating 2; the sector on the right dips below it and is rated by the other camera.

The same networks locate the weld so the camera sets its own zoom, iris and focus. A second capability detects every spatter particle on the electrode face, about 1,000 in one image, and reports a size distribution for trending. Heat and electrode identifiers come from the control system; the operator's rating triggers capture; the computed rating returns to the PLC and, with both images and both ratings, to the MES record.

Close-up of a green electrode face with dozens of bright spatter particles each boxed in red, below the dark weld bead.
FIG. 21Spatter detection on the electrode face. 1 Each particle is detected, cropped, edge-fitted and measured. 2 The weld bead above is excluded. The output is a size distribution per electrode, trended rather than judged by eye.
Screenshot of the weld-rating application with operator input fields, two camera views with overlays, left and right camera ratings of 2 and a Done status.
FIG. 22The operator workstation. 1 The operator enters a rating and submits, which triggers capture. 2 Both cameras, about 180° apart, each with the overlay. 3 The system's rating per camera. 4 Status: the result has gone to the welder PLC and the MES record. A disagreement between operator and system stops work for assessment.
Operator enters a ratingand triggers captureTwo camerascapture the weldTwo CNNs segmentweld and electrodeElectrodeedge found?No rating:welder lockednoyesGeometric rating1 to 5 from coverageRating1 to 3?Matches theoperator's rating?yesPLC permitsthe meltyesRating 4 or 5:PLC blocks the meltnoRatings disagree:stop and assessnoMES record for every resultheat and electrode ID, operator,both ratings, both images
FIG. 23Decision logic. Red: the three outcomes that stop the process: no electrode edge (welder locked), rating 4 or 5 (PLC blocks the melt), ratings disagree (stop and assess).

Impact

The rating is a reproducible measurement delivered in 5 to 10 seconds; a weld that fails cannot proceed to the furnace; every weld has a traceable record. A provisional patent application covers the method.

Limits

Agreement with an engineer was about 85% on a 27-image set mid-development, with the disagreements at rating-band boundaries, so the band thresholds are the sensitive parameter. The spatter measurement trends; it does not gate.

03 Multi-sensor fusion on one record: straightness, hot and coldMeasured billet straightness at the forge press in real time, measured by 4K video collected 24/7, hot results benchmarked against the cold laser scanner on the same billets and results on the operator's pulpit display, by combining semantic segmentation with least-squares edge slopes and fusing both instruments on one MES record.

His part: technical lead and coder for the measurement method and the camera system at the press, and for its fusion with the cold-billet laser scanner built in his organization. Period: scanner from about 2020, hot measurement from about 2022.

Problem and constraints

A forged billet has to be straight, and straightness is set while it is hot. The laser scanner is the accurate instrument but only exists after cooling. At the press the billet glows, changes color, sits in front of a cluttered background and varies in diameter and length; pixel-level edge noise is amplified by a slope; and two instruments at two stations need a shared identity.

Decisions and trade-offs

  • Measure hot, and benchmark the measurement. A camera at the press is less accurate than the laser and arrives in time to act on, so both were built and the camera result is calibrated against the laser result of the same billet.
  • Segmentation over classical edge finding. Version one found the billet with a detector and the edges classically; it broke on clutter and on changes in diameter, length and color. Version two removes the background with semantic segmentation first.
  • Filter before differentiating. Edge coordinates pass through a median filter (windows of 11, 25 and 101 samples were compared) before the slope is taken.

Solution

Wide photograph of a glowing orange billet on press tooling in a large forge building with structural steel and furnace doors behind it.
FIG. 24The press camera's view at the moment of measurement. 1 The billet, glowing and changing color as it cools. 2 The background the detector has to reject: cranes, railings, furnace doors. 3 Press tooling in the foreground.
Photograph of a hot billet with a translucent red segmentation mask over it; the background and cradles are unmasked.
FIG. 25Version two: the billet isolated by semantic segmentation before any edge is measured. 1 The mask covers the billet and nothing else. 2 The press structure behind it is excluded. 3 So are the cradles it rests on.
Hot · at the forge pressCold · after forgingForge pressHot billet4K camera,24/7Press PLCpress signalsVision workstationsegmentation + geometryBillet record in the MESkeyed by billet IDScanner controllerlaser profilesID · resultID · resultCold result calibrates the hot measurementscanLaser carriageCold billetCradle
FIG. 26Two stations, one record. Red arrow: the closed loop in which the cold laser result calibrates the hot camera measurement.

Straightness is a change in diameter over a change in length, the slope of an edge. Each edge is a least-squares line through its pixel coordinates, and the difference between the top and bottom slopes is the average straightness, compared with the 1 mm/m criterion; for billet i the hot and cold results sit under one MES identifier and are compared:

Two-panel figure: the billet in the press camera frame with four red corner points, and below it slope curves along the billet with section means and a threshold line.
FIG. 27Measurement output on one billet. 1 One of the four corner points that bound the top and bottom edges. 2 The mean slope difference in each of six sections. 3 The threshold line. 4 The difference curve (top slope minus bottom slope) along the billet.
Slope of one edgeθP1 (x1, y1)P2 (x2, y2)Δ lengthΔ dia.slope = Δ dia. / Δ length = tan θTop and bottom edges, by sectionS1S2S3S4S5S6top edge: least-squares slope_topbottom edge: least-squares slope_bottomslope_top − slope_bottomaverage straightness = slope_top − slope_bottom, criterion 1 mm/m
FIG. 28The measurement defined. Red, left: the rise of one edge over its length. Red, right: the difference between the top-edge and bottom-edge slopes, evaluated in sections.
Photograph of an aluminum extrusion rail with a sensor carriage above a round billet resting on yellow cradles.
FIG. 29The cold-billet laser scanner. 1 The laser carriage on 2 the extrusion rail that runs the length of the billet. 3 The billet on 4 its cradles. This is the reference the hot measurement is calibrated against.

Impact

Straightness is measured at the press in real time from video collected around the clock and goes to the operator's pulpit display; the hot measurement is benchmarked against the laser measurement of the same billet on one MES record. It is stage three of the platform roadmap in practice: from seeing a billet better to measuring it against a second reference.

Limits

The hot-versus-cold comparison was made on specific heats, not every billet. The camera result's engineering meaning comes from that calibration, not from pixel slopes alone.

04 Laser metrology and in-process control (US 2025/0277286 A1)Put inspection on a measured footing, measured by about 80% of final-melt crucibles at one plant scanned, an LMPC 2024 Most Innovative Work Award and a pending patent on shelf detection, by building a laser-and-camera crucible scanner and the in-process measurements that feed arc control.

His part: coded the crucible scanner and the shelf-detection control; led the rest with his organization.

Crucible inspection scanner. A telescoping laser-and-camera head scans each crucible in height and angle (six lasers at 60°, 50,000+ points per second), measures circularity, taper, spatter high points and arc gouges, flags defects with an ML model and tracks each crucible's health by ID.

Photograph of a tall scanner rig in a workshop: a vertical telescoping mast above a steel base frame, with a crucible section on a stand beside it.
FIG. 30The scanner prototype on the floor. 1 The telescoping mast that carries the laser-and-camera head down the crucible. 2 The head at the bottom of the mast. 3 The base frame that straddles the crucible. 4 A crucible section used for trials.
Engineering drawing sheet with plan, elevation and isometric views of a two-position crucible scanner with dimensions.
FIG. 31The production scanner as drawn for the site. 1 Plan with two crucible positions. 2 Elevation with clearances and maximum load height. 3 The isometric used for the installation package.
Three panels: a measured radius profile with a step, the slope-change curve with a peak, and a camera image with a thin vertical laser line.
FIG. 32One scan line from the head. 1 Laser-measured radius against height, with a step where the wall surface changes. 2 Its derivative, the slope change, which is what flags a spatter high point or a gouge. 3 The camera frame with the laser line it was measured from.
Two plots: a color map of crucible diameter over height and angle, and a diameter-range plot by height.
FIG. 33A full scan, one crucible. 1 Diameter mapped over height and angle. 2 The diameter range by height, with the measured mean. Circularity, taper and local defects come from this record, and each crucible's history is tracked by ID.
Four photographs looking down into crucibles: two clean walls marked pass, two with spatter build-up or damage marked fail.
FIG. 34What the scanner looks for, by eye first. 1 A serviceable crucible wall. 2 Spatter build-up and 3 a damaged wall that must come out of service. The scanner turns this judgment into a measured profile.

Shelf detection and control. Spatter can build up on the crucible wall as a shelf that later breaks off. A CNN selects the region of interest, edge detection finds the crucible, shelf and electrode edges, and the shelf's share of the gap sets the coil current that steers the arc (US 2025/0277286 A1, pending):

Region of interest in the gapcrucible edgeelectrode edgegapshelf thicknessshelf edgeRegion of interestselected by a CNNEdge detectionthree edgesShelf share of gapthickness ÷ gap widthController adjusts coil currentsteers the arc to limit shelf growth
FIG. 35Shelf detection. Red: the shelf edge and its thickness, measured as a share of the gap and fed to the controller.

Also in production. Alignment and centering checks before the melt (electrode-stub concentricity, weld centering, electrode geometry, electrode-crucible centering, ram centricity); real-time electrode-crucible gap and arc monitoring; electron-beam furnace analytics (condensate, liquid level, casting rate, mold temperature); automated quality scoring of each run from its video; ingot surface rating to the plant historian; wafer flatness; nonconformance decisions from process data; laser cut-location guidance to the saw after operator confirmation; forge-press personnel safety monitoring; CT defect detection and process vision for castings divisions.

Controls, mechanical design and automation

05 Closed-loop process control from vision (US 12,332,611 B2)Automated end-of-melt power cutback in production for 3+ years, measured by 85%+ lower prediction error within four months and U.S. Patent 12,332,611, by replacing a licensed, hard-wired first-generation system with an in-house CNN-plus-regression trigger connected to the furnace at two interface points.

His part: led the second generation from research to production and coded it; co-inventor on the granted patent; co-author of the LMPC 2022 Best Paper. Period: first in-house version 2016, production-ready 2019, patent granted 2025. US 12,332,611 B2 · LMPC 2022 abstract.

Problem and constraints

Screenshot of a LabVIEW application with a furnace image, panels of threshold parameters and a brightness profile plot.
FIG. 36The first generation, built with the University of Birmingham: a licensed LabVIEW application. 1 The camera image with hand-set pin thresholds. 2 Dozens of tuning parameters, adjustable only with the furnace down. 3 The brightness profile the pins were found on. Replacing this is what the rebuild was for.

Each electrode carries marker pins at a set distance from its end; when the melt reaches them, power should begin to ramp down, and too early or too late hurts ingot quality. The first-generation system, started in 2014 with the University of Birmingham, was hard-wired into the control room, ran on licensed software and needed specialists for changes that could only be made with the furnace down. The replacement acts on process power, so it must fail safe, cope with visibility that varies between furnaces and over a melt, and be maintainable by plant engineers.

Decisions and trade-offs

  • Rebuild in-house rather than keep paying for the first generation. The rebuild took from 2016 to production-ready code in 2019; in exchange the plant owns the system and the framework carried a dozen later systems.
  • Two independent triggers. Pin counting from video and a regression on process data can each start the cutback; together they cover frames where visibility fails and runs where the process drifts.
  • Operator first, automatic second. A fully automatic cutback was possible. The chosen design alarms, offers decision support, and acts only if no one responds.

Solution

A convolutional network detects and counts the marker pins in each frame. In parallel, a regression on process parameters pj (electrode weight, current, voltage) predicts the ram position at which cutback is due; either condition starts the power-adjustment phase:

Video frame of the bright electrode edge with three marker pins boxed in green and text overlays: frame number, pins 3, predicted and live positions.
FIG. 37The in-house system's live frame. 1 Frame number and the pin count from the network. 2 The pins found and boxed on the electrode. 3 Predicted cutback position from the regression against the live position: the second trigger.
Dim red video frame of the electrode edge with pins faintly visible in a green box and the overlay reading pins N/A.
FIG. 38When visibility fails. 1 The pin count reads N/A. 2 The pins are barely visible in the glow. The regression trigger on process data carries the cutback, which is why the system has two.
First generation, started 2014Camerafurnace viewCustom capturelicensed hardwareVision softwarelicensed, specialists onlyFurnace controlscontrol-room consolehard-wiredpin positions adjusted by hand · changes only during downtimeIn-house versiontwo interface pointsCamerastandard modelRack serverCNN + regression, PythonFurnace PLCramp-down requestOperatoralarm, decision supportOpen video historianfiberalarmoperator confirms; automatic if no response
FIG. 39What changed between generations. Red arrow: the entire connection to the process controls is two interface points, in place of a hard-wired, specialist-maintained link.

The rollout replaced bespoke parts with standard cameras, fiber to every furnace, rack servers, PLC-to-PC cards and video storage under IT policy agreed in advance. A decentralized video historian gave engineers direct access to footage and became the training source for later models.

Impact

Prediction error fell 85% or more within four months of development. The system has run in production for more than three years, won the LMPC 2022 Best Paper Award, and its framework (standard hardware, two interface points, tools plant staff can troubleshoot) carried gap measurement, shelf detection, arc monitoring and debris detection, and crossed to an electron-beam furnace, a forge press and a castings division.

Limits

Pin visibility varies between furnaces, cameras and the course of a melt; the regression trigger exists because the vision trigger cannot be the only one. The 85%+ figure is a development-period measurement over four months.

06 Electromagnetic actuation and control (US 2025/0119992 A1)Closed a ~2% ingot-yield gap between two plants, measured by 19+ ingots at the top surface rating, by designing a twin-coil arc-steering retrofit with three-phase PWM drives, PLC control and camera-based verification of arc rotation.

His part: technical lead for the electromagnetic design, controls and verification, coded the controls; co-inventor. US 2025/0119992 A1, pending.

Problem and constraints

Surface imperfections on an ingot must be ground away, which costs yield. One plant controlled this with an external magnetic field that moves the arc during melting; its ingots yielded up to 2% more. The sister plant's furnaces could not accept that design without structural changes, the field has to follow the arc as it climbs the crucible, and the top of the melt is out of a moving coil's reach.

Decisions and trade-offs

  • Retrofit with two coils instead of rebuilding the furnace. A moving coil on a carriage that tracks the arc zone, plus a stationary coil for the top of the melt, fits the existing furnace at the cost of a motion system and a second drive.
  • A bench rig before the furnace. Coils, drives and PLC were built up on a test setup so controls could be tuned without furnace time.
  • Verify the rotation. Image analysis of the furnace camera tracks the arc's bright region and confirms one intensity peak per revolution, so the field's effect is measured rather than inferred from coil current.

Solution

Each coil has three pairs of opposing segments driven by sinusoidal currents 120° apart, which sum to a field of constant magnitude rotating at the electrical frequency; the carriage follows ingot height, derived from ingot weight. COMSOL electromagnetic and CFD models set the coil geometry; melt trials set current and voltage, rotation speed and reversal frequency, and coil position relative to the arc gap.

Furnace sectioncarriage riseswith the arcStationary coilElectrodeMoving coilArc zoneMolten poolSolid ingotWater-cooled crucibleTop view of one coilAC′BA′CB′rotatingfieldThree opposingsegment pairsSegment currents, 120° apartABC
FIG. 40Twin-coil arc steering. Red, left: the moving coil, which rises with the arc zone. Red, right: the rotating field from three segment currents 120° apart. Simplified from US 2025/0119992 A1.
Furnace controllerrecipe, ingot weightArc-steering PLCfield and carriage commandsMoving-coil drivethree PWM H-bridge phasesStationary-coil drivethree PWM H-bridge phasesMoving coilpairs A–A′, B–B′, C–C′Stationary coilsame pairs, driven in phase3-phasecoil current and resistance tapsCarriage motor drivelead screwCoil carriagetracks ingot heightEncoder and limit switchesposition, travel limitsposition feedback
FIG. 41Control architecture. Red arrows: the three-phase currents to each coil. Dashed lines are feedback to the PLC.
CAD side view of the twin-coil arc-steering assembly: a stationary upper coil and a moving lower coil on a carriage around the crucible.
FIG. 42Twin-coil assembly in CAD: stationary coil above, moving coil on its carriage below.
Photograph of a crucible in a furnace pit with a carriage plate part way down, vertical guide rods and a drive at the base.
FIG. 43The retrofit in the furnace pit. 1 The crucible, around which both coils sit. 2 The moving-coil carriage, which rises as the ingot grows. 3 A guide rod the carriage rides on. 4 The carriage drive at the base.
Photograph of an open control cabinet with a PLC rack at the top, terminal rows, a drive, driver boards and two power supplies at the bottom.
FIG. 44The arc-steering control cabinet. 1 The PLC rack that runs the field program and the carriage. 2 Terminal and relay rows to the furnace controls. 3 The carriage drive. 4 Coil drive boards. 5 DC power supplies feeding the H-bridges.
Two plots: process data over a melt including a rising ingot-height curve, and arc rotation in revolutions per minute holding near 3 with excursions at start and end.
FIG. 45Verification from the process record. 1 Ingot height over the melt, from which the carriage position is derived. 2 Arc rotation in revolutions per minute, measured from the furnace camera, holding at the commanded speed through the melt.

Impact

The system runs in production on the retrofitted furnace and closed the roughly 2% yield gap between the two plants; 19+ ingots have come out at the top surface rating, smooth, with no porosity, blisters or rundown, consistent from melt to melt. The bench rig lets controls changes be tuned away from the furnace.

Photograph of the end of a titanium ingot with a rough, flaking surface and heavy crown material around the rim.
FIG. 46Before: an ingot from the sister plant. 1 Rough, flaking surface and 2 heavy crown material around the rim, which has to be ground away. That is where the ~2% yield goes.
Photograph of a titanium ingot on stands with a smooth, uniform surface.
FIG. 47After: an ingot melted with the twin-coil system. 1 A smooth surface along the full length, no porosity, blisters or rundown. 19+ ingots have come out at this rating.

Limits

The surface-rating result stands on 19+ ingots from the retrofitted furnace; the yield comparison is between two plants with other differences besides the field. The current equation states the design intent, not a measured field map.

Patents and papers

Patents are assigned to a Precision Castparts subsidiary. LMPC is the Liquid Metal Processing & Casting Conference, where Herscovici also served on the committee. He spoke and led workshops on machine learning and computer vision at PCC's technical conferences each year.

Background

Roles at Precision Castparts Corp.

  • 2020–2026Head of Manufacturing Technology and Principal Engineer, R&D
  • 2017–2020Senior Manufacturing Process Technology Engineer
  • 2015–2017Research and Data Analytics Engineer, R&D
  • 2014–2015Manufacturing Process Technology Engineer, R&D

Education

M.S. Mechanical Engineering · B.S. Mechanical Engineering · B.S. Computer Science

University of Nevada, Las Vegas

Working tools: Python, PyTorch, TensorFlow, OpenCV, scikit-learn, ONNX, TensorRT; Databricks, Spark, MLflow, Azure; Allen-Bradley PLCs over OPC-UA, Ignition HMI, AVEVA MES; SolidWorks, COMSOL, MATLAB and Simulink. C/C++ and CUDA shipped in production earlier in his tenure; recent work is Python.

Notes. Patents, the conference paper and the plant groundbreaking are public record and linked where they appear. Production figures are company records, stated as the floor of any range. Proprietary detail is left out: no site codes, no process recipes, no code. Figures are simplified sketches, not to scale. Contact details are on the CV.