Engineering portfolio

Nir Herscovici

Engineering leadership across software, sensing and control

Nir Herscovici is a principal engineer and R&D leader with more than a decade of experience developing integrated hardware and software for aerospace manufacturing. His background in mechanical engineering and computer science connects algorithm development with instrumentation, electromechanical design and control.

At Precision Castparts, he built and led a 22-person multidisciplinary R&D organization, managed a seven-figure R&D budget and owned the $5M+ technology scope for a new manufacturing facility. His organization deployed 15+ systems across ten facilities in three countries. He also served as the customer-facing technical lead on an investigation involving nine aerospace OEMs.

His work translates neural-network assessments into operator guidance and equipment interlocks, with defined responses when measurements are unavailable or inconsistent. Electromechanical actuation and feedback control have delivered yield improvements of up to 5%, while equipment-health monitoring supports preventive maintenance. He combines laser metrology with computer vision to measure part geometry, verify alignment and identify defects, giving engineering and operations teams repeatable evidence for critical decisions.

Across participating plants, broader program outcomes included 25%+ fewer recordable safety incidents, 30%+ reductions in scrap and defects, and 10%+ improvements in uptime and throughput.

His technical record includes patented and patent-pending technologies, publications, conference presentations and workshops. His work received LMPC’s 2022 Best Paper and 2024 Most Innovative Work awards.

  • 22person R&D team
    built and led
  • 15+systems in continuous
    production operation
  • 10facilities across
    three countries
  • $5M+technology scope
    owned and delivered

Head of Manufacturing Technology & Principal Engineer, R&D
Precision Castparts Corp. · 2020–2026

Engineering leadership

Technical direction, multidisciplinary teams, investment and program delivery, grounded in direct engineering experience.

Explore leadership

Software & computing

Python applications, machine learning, geometric algorithms and data processing, from analytical methods to deployed software.

Explore software

Hardware & control

Instrumentation, acquisition interfaces, PLC integration and electromechanical design, supported by physical verification.

Explore integration

Technical direction & execution

Leadership & delivery

Herscovici set technical roadmaps, managed a seven-figure R&D budget and secured more than $2M in OEM and cross-division funding. He recruited engineers, advised graduate researchers and developed a multidisciplinary organization while remaining involved in software, controls and technical verification.

Team development, program ownership and results

Build engineering capability

He led ten engineers, five graduate researchers and seven technicians, hiring or recruiting seven of the engineers and advising all five researchers. Every engineer earned a promotion; one technician advanced to superintendent. Twelve of fifteen interns joined full time, and the internship structure was adopted company-wide.

Own delivery and investment

He managed a seven-figure R&D budget, presented roadmaps to the COO and secured CEO approval to expand the organization. Capital proposals typically targeted payback below twelve months. He also secured more than $2M in OEM and cross-division funding.

Coordinate complex integration

At a $500M greenfield facility, he owned the $5M+ technology scope, coordinating contractors, the equipment supplier and delivery milestones. Instrumentation for seven furnaces was part of the program from site inception.

Lead customer engineering

He served as technical lead to nine aerospace OEMs, including Pratt & Whitney, GE Aerospace, Rolls-Royce and Lockheed Martin, during a material-defect investigation. He also led image analysis in a DARPA-funded consortium involving four OEMs.

Reported operational results

At participating plants, program outcomes included 30%+ reductions in scrap and defects, 25%+ fewer recordable safety incidents, and 10%+ improvements in uptime and throughput.

Delivery discipline

Business cases and trade studies informed investment. Stage-gate reviews governed progression from pilot to production; weekly schedules and executive roadmaps made progress and dependencies visible. Standard hardware, in-house software and documented procedures established ongoing ownership within the operating teams.

Career progression and milestones
RolesHead of Manufacturing Technology and Principal Engineer, R&DSenior Manufacturing Process Technology EngineerResearch and Data Analytics Engineer, R&DManufacturing Process Technology Engineer, R&D2014201620182020202220242026Builds internshipprogramTalk: vision forprocess controlFirst directengineering hiresTalks: equipment health,ML and visionTalk: greenfieldtechnology roadmapTalk: company-widetechnology roadmapMilestonesVision program starts,Univ. of BirminghamFirst in-houseprocess visionNeural-networkdevelopment beginsHead of a 22-personR&D organizationProvisional application,AI process controlLMPC BestPaper AwardSelected to leadgreenfield technology$500M greenfieldplant breaks groundLMPC Most InnovativeWork AwardU.S. patent granted;two more publishedInspection filing;conference paper
  1. 2014Manufacturing Process Technology Engineer, R&D. Vision program starts, Univ. of Birmingham
  2. 2015Research and Data Analytics Engineer, R&D
  3. 2016First in-house process vision
  4. 2017Senior Manufacturing Process Technology Engineer. Builds internship program
  5. 2018Neural-network development begins
  6. 2019Talk: vision for process control
  7. 2020Head of Manufacturing Technology and Principal Engineer, R&D. Head of a 22-person R&D organization. First direct engineering hires
  8. 2021Provisional application, AI process control. Talks: equipment health, ML and vision
  9. 2022LMPC Best Paper Award. Selected to lead greenfield technology
  10. 2023$500M greenfield plant breaks ground. Talk: greenfield technology roadmap
  11. 2024LMPC Most Innovative Work Award
  12. 2025U.S. patent granted; two more published. Talk: company-wide technology roadmap
  13. 2026Inspection filing; conference paper
Career progression at Precision Castparts, 2014–2026.

Selected work

Technical case studies

Six studies make his contribution explicit: the problem, what he designed or implemented, the engineering decisions and the resulting evidence.

01 Detection and Tracking with Sparse DataNeural networks screening 30,000 hours of video for rare critical events95% fewer false positives · manual review cut 100× · resolved for all nine OEMs

Problem

A rare contamination event, a few pixels wide and present for only a handful of frames, had to be located across more than 30,000 hours of production video. Glare, flicker and sensor noise produced near-identical signatures, and only 73 confirmed examples existed to learn from. Manual review could not scale, and every finding had to withstand audit by nine aerospace customers.

Solution

A high-recall neural detector screens each frame, a motion tracker confirms genuine events by their physical trajectory, and a data-quality monitor flags footage that falls outside the training distribution. Processing ran on a distributed platform alongside time-aligned process telemetry, so every confirmed event is preserved with its frames, trajectory and process context. The output is a reviewable chain of evidence rather than a list of alarms.

Scope of work

Role
Designed and implemented the detection and tracking pipeline; customer-facing technical lead for the investigation.
Team
More than 50 contributors across sites and divisions, about 30 of them engineers; weekly review calls with customer executives and specialists.
Data
1,132 production runs, 2,200 videos, 30,000+ hours, about 4 TB, time-aligned with process telemetry.
Customers
Nine aerospace OEMs.
Impact
95% fewer false positives at 95%+ confirmed-event recall; manual review cut 100×; projected review cost down about 80%; the defect was resolved for all nine OEMs.

Technical details

Capabilities: Computer vision · Kalman tracking · distribution analysis · distributed processing
Tools: Python, PyTorch, OpenCV, scikit-learn, Databricks, Spark

Engineering challenge

Contamination events involved targets only a few pixels across, visible for a handful of frames. Glare, flicker, and sensor noise produced similar observations. Only 73 confirmed examples were available, camera conditions varied, and each reported event required evidence suitable for customer review.

Architecture and decisions

  • Preserve small targets. A tiled detector with a high-resolution detection head retained image detail. Restricting analysis to the physically relevant region reduced irrelevant candidates.
  • Combine appearance with motion. A high-recall CNN generated candidates; a random forest, spatial-temporal clustering, and constant-acceleration Kalman tracking tested appearance, continuity, and physical plausibility.
  • Evaluate the data as well as the model. Distribution analysis exposed brighter, sharper training images than field recordings. Dataset revision addressed that mismatch, while a monitor flagged unfamiliar video conditions.
Contamination-event detection architecture: synchronized video and telemetry, neural detection, appearance validation, temporal validation and engineering review.
FIG. 1 Contamination-event detection architecture. Synchronized video and telemetry support distributed analysis; candidate validation produces a traceable record for engineering review. Inspection geometry is illustrative. AI-generated conceptual schematic.

Databricks supported distributed video processing with time-aligned controller telemetry. Timestamps, frames, trajectories, and engineering assessments remained attached to each event, preserving traceability through the review.

Evidence and outcome

The investigation supported resolution of the defect for nine OEMs. Reported evaluation showed 95% fewer false positives at 95%+ confirmed-event recall, reducing manual review 100× and projected review costs by approximately 80%.

All 73 confirmed events were detected, but those examples also informed tracker and clustering calibration. Recall therefore describes confirmed-event evaluation. The distribution monitor identifies unfamiliar conditions. Dataset pruning and revised training coverage improved detector performance.

Algorithm detail and detection constraints

The image-quality monitor compares brightness, sharpness, and region size with the training distribution using median- and IQR-based distances. Feature-space analysis guided removal of near-duplicate bright, sharp images and addition of representative field conditions.

A temporal mean, Sobel and Canny edges, morphology, and ellipse fitting locate the region of interest. The random forest evaluates 2,242 texture, gradient, and shape features. Statistical gating rejects detections inconsistent with the predicted track.

Four detection and validation steps: visible-region masking, tiled CNN and random-forest validation, spatial-temporal clustering, and Kalman association with distant observations excluded from the track.
FIG. 2 Detection and validation methods. Region masking, neural detection and appearance filtering retain candidates; clustering consolidates repeated observations, and statistical motion gating establishes a consistent track. The distant observation sequence is excluded from that track. AI-generated conceptual schematic.
Deployment-quality assessment, representative training-data revision and comparison of reference and revised detectors on identical evidence, followed by reported system results.
FIG. 3 Deployment robustness and model verification. Video conditions are assessed against the training distribution; pruning and representative examples refine the dataset. Reference and revised detectors are compared on common evidence. Reported results describe known-event evaluation. AI-generated conceptual schematic.
02 Perception, Measurement, and Control InterlocksNeural segmentation tracing part geometry for complianceRating in 5–10 s with an image record · among the systems behind 30%+ less scrap

Problem

Ahead of a critical melting step, weld quality rested on an operator’s eye. Ratings drifted between shifts, no record was kept, and a weld that should have failed could proceed into production. The replacement had to return a repeatable result within seconds under variable lighting and clutter, with a defined response whenever a measurement was unavailable or disputed.

Solution

Two cameras, mounted roughly opposite each other, image the weld; neural networks trace the weld and part boundaries, and geometric analysis converts those boundaries into an explicit five-level rating within seconds. Control authority stays with the plant controller: a missing boundary locks the welder, a failed rating holds the next step, and disagreement with the operator triggers review. Every assessment is retained with its image and part record.

Scope of work

Role
Led and implemented the system; co-inventor on its provisional patent; co-author of the related conference paper.
Timeline
2025–2026.
Integration
Plant controller (PLC) interlock and manufacturing record (MES); two cameras and lighting mounted on the existing equipment.
Impact
Rating in 5–10 seconds with an image record for every assessment; one of the inspection and monitoring systems credited with cutting scrap and defects by more than 30%.
Validation
About 85% agreement with an engineer across 27 images mid-development, with differences near the rating boundaries.

Technical details

Capabilities: Semantic segmentation · geometric measurement · decision logic · control integration
Tools: Python, PyTorch, OpenCV, PLCs, MES

Engineering challenge

Operators assessed electrode welds visually before melting, with ratings varying between shifts. The replacement needed a repeatable result within seconds despite clutter and variable lighting, a defined response to unavailable measurements, and traceability to the part and operator.

Architecture and decisions

  • Separate recognition from assessment. Two segmentation networks identify the weld and electrode boundaries. An ellipse establishes the centre, and radial measurements convert those boundaries into an explicit five-level rating.
  • Improve observability. Cameras approximately 180° apart assess complementary sectors, reducing dependence on a single viewpoint.
  • Allocate control authority explicitly. The PLC retains the interlock. A missing electrode boundary produces no rating and locks the welder; ratings 4 and 5 block the melt; disagreement with an otherwise acceptable operator rating triggers assessment.

Engineering sketch

AI-generated engineering sketch preserving the asymmetric weld-pool boundary and its position relative to the outer electrode edge.

Assessment sketch

AI-generated engineering sketch matching the raw weld geometry, with the weld contour and elliptical rating bands.
FIG. 4 Matched engineering sketches of the weld and its assessment. The red weld contour and elliptical rating bands explain the illustrative rating 3. AI-generated conceptual illustration.

Control-system identifiers associate both images and ratings with the manufacturing record. The operator’s assessment triggers capture, and the computed result passes to the PLC and manufacturing execution system (MES).

Operator enters a ratingand triggers captureTwo camerascapture the weldTwo CNNs segmentweld and barBaredge found?No rating:welder lockednoyesGeometric rating1 to 5 from coverageRating1 to 3?Matches theoperator's rating?yesPLC permitsthe runyesRating 4 or 5:PLC blocks the runnoRatings disagree:stop and assessnoMES record for every resultbatch and part ID, operator,both ratings, both images
FIG. 5 Decision logic. Red: missing boundary, unacceptable rating, and operator disagreement each lead to a defined stop or review condition.

Evidence and outcome

The system returns a geometric assessment in approximately 5–10 seconds and applies explicit acceptance logic with an image record for every assessment. Mid-development agreement with an engineer was approximately 85% across 27 images. Differences clustered near rating boundaries, making threshold calibration a material part of validation.

Measurement geometry and rating thresholds

The ratio shown as “coverage” in the development figures is radial extent, rw(θ)/re(θ). It compares weld and electrode boundaries along each assessed ray. The maximum ratio determines the rating:

Image geometry2345θCr_e(θ): bar edgeedge on this rayr_w(θ): furthest weldpoint on this rayHolder (hatched)Weld contour(segmentation CNN)Best-fit ellipse Eof the bar faceTop sector ignored:rated by the second cameraDashed ellipses: rating bands 2 to 5Coverage around the barignored sector20.6030.7540.9050.97rating0°90°180°270°360°angle θcov(θ)peak: rating 4
FIG. 6 Geometric rating. Red, left: the weld boundary along one ray. Red, right: the corresponding ratio peak crosses the rating-4 threshold.
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. 7 Radial extent ratio by angle. 1 Peak above the rating-4 threshold. 2 Rating-2 threshold; the opposite camera assesses the sector outside this view.

Installation, operator interface, and particle analytics

StationSystemCamera 1Camera 2LightLightHolderWeld pool:what is ratedBarWelderchambercameras about 180° apartGPU workstationweld CNN + bar CNNgeometric rating, 5–10 sOperator interfaceoperator rating, result,both camera viewsRecorder24/7 videoWelder PLCbatch and part IDsinterlockMES and historianrating record per batchHD video (SDI)IDs in,rating outInterlock: locks the welder,blocks the run
FIG. 8 Inspection architecture. Red, left: assessed weld region. Red arrow: PLC interlock connecting the assessment to equipment operation.
Engineering sketch of the welder frame, camera and illumination mounts, and lower junction box.
FIG. 9 Engineering sketch of the installation design: cameras and illumination on the welder frame, with a junction box at the base. AI-generated conceptual illustration.

The segmentation outputs also support camera zoom, iris, and focus adjustment. A separate particle-analysis pipeline measures spatter-size distributions for process trending; these measurements do not control the acceptance interlock.

Full-view sketch

Engineering sketch of the electrode, stub and weld bead, with the region used for spatter analysis marked.

Enlarged, with detections

Matched engineering closeup with loose spatter particles boxed and the fused weld bead excluded.
FIG. 10 Matched engineering sketches of the full weld view and enlarged particle detections. Loose spatter is boxed; the fused weld bead is excluded from the count. AI-generated conceptual illustration.
Engineering sketch workflow using one repeated weld scene and one repeated spatter particle, from detection through contour extraction and geometric fitting.
FIG. 11 Illustrative particle-measurement workflow using the same weld scene: region selection and detection, then one particle through grayscale extraction, edge analysis and geometric fitting. The measurements support per-part size distributions. AI-generated conceptual illustration.
Conceptual weld-rating interface sketch with operator controls, two camera views, rating 2 from both cameras, and PLC / MES record status.
FIG. 12 Engineering sketch of the operator interface: operator rating and capture trigger, two annotated camera views, computed ratings and PLC / manufacturing-record status. The rating-2 example is separate from FIG. 4. AI-generated conceptual illustration.
03 Vision Metrology and Cross-Sensor ValidationSensor fusion for robust process measurementGeometry feedback while correction is still possible · laser cross-check on every part

Problem

A forged part can be corrected only while it remains hot, yet the accurate laser measurement was taken after cooling. At the press, operators had no dimensional feedback during the window in which correction was still possible. The glowing surface changes appearance against a cluttered background, so any camera measurement had to be both timely and trustworthy.

Solution

A press-mounted 4K camera measures the glowing part in real time, with neural segmentation isolating the part from tooling before any geometry is fitted. A laser station measures the part again once cold, and both results are linked to the same part record, so each camera reading is validated against the laser reference. Operators act while correction is possible, and the independent measurement confirms the call afterwards.

Scope of work

Role
Led and implemented the press-camera system, its measurement method and its integration with the laser scanner built by his organization.
Timeline
Laser scanner from about 2020; hot-part camera measurement from about 2022.
Integration
PLC-triggered capture; the manufacturing record links camera and laser results by part identifier.
Impact
Geometric feedback during forging, with an independent cold reference for every measured part; the laser station also guides cut locations, with operator confirmation.

Technical details

Capabilities: Vision metrology · segmentation · noise filtering · cross-sensor validation
Tools: Python, OpenCV, PyTorch, laser triangulation

Engineering challenge

Dimensional feedback was needed while a forged billet remained hot enough to correct. The more accurate laser reference measured parts after cooling. At the press, the glowing surface changed appearance against a cluttered background, part dimensions varied, and pixel-level boundary noise affected geometric estimates.

Architecture and decisions

  • Match measurement to the decision. A continuously operating 4K camera supplied timely geometry at the press; downstream laser profiles provided an independent reference for evaluating that measurement.
  • Isolate the object before measuring it. Semantic segmentation replaced an earlier detector-and-edge approach, separating the billet from tooling and background structures.
  • Control noise before fitting geometry. Median filtering smoothed boundary coordinates before least-squares fitting. Development compared 11-, 25-, and 101-sample filter windows.
AI-generated conceptual illustration of a glowing hot billet on forge supports.
FIG. 13 Engineering sketch of the operating scene. The glowing billet, foreground tooling and background structure show the conditions faced by camera measurement. AI-generated conceptual illustration.
Engineering sketch of the same billet scene with a translucent segmentation mask excluding cradles and press structure.
FIG. 14 Matched segmentation sketch of the billet in FIG. 13. The billet mask excludes press structure and foreground cradles before measurement. AI-generated conceptual illustration.

PLC signals triggered measurement. A common part identifier linked camera observations and subsequent laser scans in the MES, supporting comparison between instruments. Sectional upper- and lower-edge fits characterized the part’s profile; their slope difference measured longitudinal diameter variation.

Hot · at the forge pressCold · after forgingForge pressHot bar4K camera,24/7Press PLCpress signalsVision workstationsegmentation + geometryPart record in the MESkeyed by part IDScanner controllerlaser profilesID · resultID · resultCold result calibrates the hot measurementscanLaser carriageCold barCradle
FIG. 15 Measurement traceability. Hot-camera and cold-laser results share a part identifier. Red arrow: the independent laser reference supports comparison and calibration.

Evidence and outcome

Operators received geometric feedback during forging, and linked records supported comparison with the cold reference. The laser station also supplied height profiles and cut-location guidance, with operator confirmation before transmission of a cut instruction.

Validation covered matched hot and cold measurements from selected production batches, so the calibration evidence is limited to those evaluated parts.

Measurement model and laser reference

Upper- and lower-edge coordinates are fitted by section using least squares. Their slope difference describes longitudinal diameter variation. The original development notation and 1 mm/m geometric criterion are retained below; they do not establish conventional centreline straightness accuracy.

Engineering sketch of the billet processing workflow, excluding tooling occlusions before fitting upper and lower edge slopes.
FIG. 16 Illustrative measurement pipeline using the same billet: region selection, enhancement, edge extraction and sectional line fitting. Cradle occlusions are excluded; upper-minus-lower slope characterizes longitudinal diameter variation. AI-generated conceptual illustration.
Slope of one edgeθP1 (x1, y1)P2 (x2, y2)Δ lengthΔ positionslope = Δ position / Δ length = tan θTop and bottom edges, by sectionS1S2S3S4S5S6top edge: least-squares slope_topbottom edge: least-squares slope_bottomslope_top − slope_bottomprofile metric = slope_top − slope_bottom, criterion 1 mm/m
FIG. 17 Edge-profile model. Red, left: position change along an edge. Red, right: the difference between fitted upper- and lower-edge slopes.
Engineering sketch of orthogonal laser profiling over a roller conveyor and a scan, review, cut-position selection, operator confirmation and saw-instruction workflow.
FIG. 18 Engineering sketch of the laser reference and operator workflow: orthogonal profiles are aligned by part position, the taper location informs the recommended cut, and the operator confirms before the cut instruction is sent. AI-generated conceptual illustration.
04 3D Inspection and Condition AssessmentSensor fusion for equipment health assessmentMeasured condition history for every asset · LMPC 2024 Most Innovative Work Award

Problem

Reusable, high-value process vessels wear, sustain damage and accumulate deposits over service. Visual inspection could neither measure the wall nor show how any vessel had changed across its life, so repair and retirement decisions rested on judgement. The internal geometry had to be measured directly, and every measurement tied to the specific asset.

Solution

A telescoping scanner carrying six lasers and cameras maps the internal surface at more than 50,000 points per second. Software computes circularity, taper and local defects, a model flags surface damage, and each result is tied to the vessel’s identity so a condition history accumulates with every inspection. Repair and retirement decisions now rest on measured data instead of visual impression.

Scope of work

Role
Implemented the inspection software and the related vision-based control; led the wider program through his organization.
Scale
About 80% of final-melt vessels at one plant; more than 50,000 measurement points per second.
Impact
A measured condition history per asset, supporting repair, retirement and maintenance decisions instead of visual judgement.
Recognition
LMPC 2024 Most Innovative Work Award.

Technical details

Capabilities: laser metrology · geometric analysis · condition monitoring · sensor integration
Tools: Python, OpenCV, laser systems, SolidWorks

Engineering challenge

Visual inspection provided limited dimensional evidence of wear, local damage and accumulated material. The system needed to measure the internal surface, identify local features and retain a condition history for each asset.

Architecture and decisions

  • Capture the geometry directly. A telescoping head carried six lasers at 60° intervals and cameras, acquiring more than 50,000 points per second across height and angle.
  • Separate measurement from classification. Geometric analysis calculated circularity, taper and local protrusions or gouges. An ML model flagged surface defects for assessment.
  • Preserve asset identity. Inspection records linked measured profiles and condition observations to a crucible identifier, supporting comparison over its service life.
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. 19 One scan line. 1 Laser-measured radius against height, with a step where the wall changes. 2 Its derivative, which flags spatter high points and gouges. 3 The camera frame with the laser line.
Two plots: a color map of crucible diameter over height and angle, and a diameter-range plot by height.
FIG. 20 A full scan of one mold. 1 Diameter mapped by height and angle. 2 Diameter range by height, with the mean. Circularity, taper and local defects come from this record.

Evidence and outcome

At one plant, inspection covered approximately 80% of final-melt crucibles. The work received the LMPC 2024 Most Innovative Work Award. Coverage describes adoption within that population; acquisition rate describes data throughput, rather than dimensional accuracy.

Related application: geometric measurement for control

Material can accumulate on a crucible wall as a shelf that may later detach. A CNN identifies the region of interest; edge detection locates the crucible, shelf and electrode boundaries.

The mean wall-to-shelf distance, normalized by the wall-to-electrode gap, defines a geometric control input. That measurement informs coil-current adjustments for arc steering. This connects a learned region detector with an explicit measurement and physical actuation.

Region of interest in the gapmold wallbar edgegapbuildup thicknessbuildup edgeRegion of interestselected by a CNNEdge detectionthree edgesBuildup share of gapthickness ÷ gap widthController adjusts coil currentsteers the arc to limit buildup
FIG. 21 Wall-buildup detection. Red: the buildup edge and its thickness as a share of the gap, fed to the controller.

Related method: US 2025/0277286 A1.

Scanner hardware and representative surface conditions

AI-generated engineering sketch of the scanner prototype: telescoping mast, compact scanning head, open base and trial mold section at right.
FIG. 22 Scanner prototype arrangement, redrawn as an engineering sketch. The telescoping mast carries a compact laser-and-camera head above the straddling base; the mold section at right was used for trials. AI-generated conceptual illustration.
Engineering drawing sheet with plan, elevation and isometric views of a two-position crucible scanner with dimensions.
FIG. 23 The production scanner as drawn for the site. 1 Plan with two mold positions. 2 Elevation with clearances. 3 Isometric for the installation package.
AI-generated examples of metal vessel interiors with serviceable walls, deposits and local damage.
FIG. 24 Vessel-wall conditions, redrawn as engineering sketches. Serviceable walls at left; deposits and local wear at right. These examples explain the features assessed by the scanner. AI-generated conceptual illustration.
05 Perception-Assisted Process ControlCustom neural nets for decision support in critical processesPrediction error down 85%+ · intervention within 1 s · LMPC 2022 Best Paper Award

Problem

A melting process must reduce power at a precise point near the end of each run. The visual cue that marks that point is intermittently obscured, and the incumbent system was licensed, hard-wired and dependent on outside specialists for any change. The replacement had to be as reliable as the cue allowed, integrate with the existing controls and remain maintainable by plant engineers.

Solution

A neural network watches for the visual cue while a regression model predicts the point from process data; either can alert the operator, and if no response follows, the system reduces power on its own. Standard cameras, fiber and rack servers with two defined control interfaces replaced the bespoke hardware, and the software is maintained in-house. The same architecture became the foundation for the measurement and monitoring applications that followed.

Scope of work

Role
Led the second-generation system from research to production and implemented its software; co-inventor on US 12,332,611 B2.
Timeline
Development 2016–2019; in production for more than three years since.
Impact
Prediction error down more than 85% over four months of development; automatic intervention within one second of detection; process change qualified with the OEM customer.
Recognition
LMPC 2022 Best Paper Award.

Technical details

Capabilities: CNN detection · regression · control integration · deployment and sustainment
Tools: Python, PyTorch, OpenCV, PLCs, OPC UA

Engineering challenge

Furnace power must decrease at a defined point near the end of an electrode melt. Visual marker pins indicate that point, but visibility varies. The replacement also had to integrate with existing controls and be maintainable by plant engineers.

Architecture and decisions

  • Use complementary observations. A CNN detects marker pins. A regression model independently estimates the trigger position from electrode weight, current and voltage. Each supplies a condition for requesting cutback.
  • Define the response sequence. The system alerts the operator with the available evidence. Automatic cutback follows under specified conditions if the operator does not respond.
  • Constrain integration and maintenance. Standard cameras, fiber, rack servers and two defined control interfaces replaced bespoke components. Internally maintained software and recorded video supported plant ownership.
First generation, started 2014Cameraprocess viewCustom capturelicensed hardwareVision softwarelicensed, specialists onlyProcess controlscontrol-room consolehard-wiredpin positions adjusted by hand · changes only during downtimeIn-house versiontwo interface pointsCamerastandard modelRack serverCNN + regression, PythonProcess PLCramp-down requestOperatoralarm, decision supportOpen video historianfiberalarmoperator confirms; automatic if no response
FIG. 25 What changed between generations. Red arrow: the whole connection to the process controls is two interface points, replacing a hard-wired, specialist-maintained link.

Evidence and outcome

Reported prediction error fell more than 85% over four months of development. The system operated in production for more than three years, and the related work received the LMPC 2022 Best Paper Award.

The improvement applies to that development interval. Variable pin visibility remains a constraint; the process-data channel provides a complementary basis for the control decision.

Prediction model and observed visibility conditions

Regression estimates the feed position for cutback from process parameters. The logical OR below identifies a trigger condition; operator notification and the configured response sequence still govern the resulting action.

Engineering sketch of three visible marker pins at the bar rim, a detection region, and independent predicted-trigger and live-position indicators.
FIG. 26 Engineering sketch of the in-house system with three visible marker pins. The vision cue and regression-predicted trigger position are independent inputs to the cutback decision. AI-generated conceptual illustration.
Matched pin-detection engineering sketch with the same rim geometry, barely visible pins, N/A vision result and an available process-data trigger.
FIG. 27 Matched low-visibility sketch. The pins remain present but obscured; the vision cue reads N/A while the process-data regression estimate remains available. AI-generated conceptual illustration.

Legacy system and the integration trade-off

The earlier system used licensed software and a hard-wired control connection. Parameter changes required outside specialists and production downtime. The in-house architecture reduced those dependencies and became a foundation for later measurement and monitoring applications.

Conceptual engineering sketch of the legacy LabVIEW interface with camera pin thresholds, many tuning controls and a brightness-profile panel.
FIG. 28 Engineering sketch of the first-generation licensed LabVIEW application built with the University of Birmingham. Hand-set camera thresholds, many tuning parameters and a brightness profile explain its specialist maintenance burden. AI-generated conceptual illustration.
06 Electromechanical Design and Feedback ControlElectromechanical systems that raise product yieldClosed a ~2% yield gap between two plants · 19+ products at the highest surface rating

Problem

One plant’s product carried a rougher surface and a lower yield than its sister plant. Steering the arc with a magnetic field would close the gap, but the existing equipment left no room for a single moving coil to reach the top of the melt, and adopting the sister plant’s design would have required structural changes.

Solution

Two coils, one rising with the process zone and one fixed at the top, are driven through power electronics by the plant controller, with camera analysis confirming that the arc rotates as intended. Electromagnetic and flow simulation set the coil geometry, a bench rig proved the coils, drives and control logic, and the assembly was retrofitted to production equipment. Camera-measured arc rotation, rather than coil current alone, served as proof of the intended effect.

Scope of work

Role
Led electromagnetic design, controls and verification; implemented the control software; inventor on US 2025/0119992 A1.
Work
Electromagnetic and flow simulation, bench rig, drive electronics, PLC control, camera-based verification, production retrofit.
Impact
Closed an approximately 2% yield gap between the two plants; more than 19 products at the highest surface rating.

Technical details

Capabilities: electromagnetic modeling · mechatronics · drive electronics · feedback control
Tools: COMSOL, SolidWorks, PLCs, PWM H-bridges, camera analysis

Engineering challenge

The field had to follow a rising process zone while fitting existing furnace geometry. A single moving coil could not reach the upper region, and transferring another plant’s design would have required structural changes.

Architecture and decisions

  • Split the operating envelope. A moving coil follows the arc zone; a stationary coil covers the top of the melt. The arrangement fits the existing equipment at the cost of a carriage and a second drive.
  • Connect models to hardware. COMSOL electromagnetic and CFD models informed coil geometry. A bench rig supported coil, drive and PLC development before furnace trials.
  • Verify the physical response. Camera analysis measured arc rotation. Coil current alone was not treated as proof of the intended process effect.
Cross-sectioncarriage riseswith the arcStationary coilBarMoving coilArc zoneMolten poolSolidified productWater-cooled moldTop view of one coilAC′BA′CB′rotatingfieldThree opposingsegment pairsSegment currents, 120° apartABC
FIG. 29 Twin-coil arc steering. Red, left: the moving coil rises with the arc zone. Red, right: the rotating field from three currents 120° apart. Simplified from US 2025/0119992 A1.

Three opposing segment pairs receive sinusoidal currents separated by 120°. The PLC commands the PWM H-bridges and carriage drive, using current feedback, an encoder and limit switches. Carriage position follows ingot height estimated from weight.

Process controllerrecipe, product 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 product heightEncoder and limit switchesposition, travel limitsposition feedback
FIG. 30 Control architecture. Red arrows: three-phase currents to each coil. Dashed lines: feedback to the PLC.

Evidence and outcome

Furnace trials established operating settings; video confirmed arc rotation. Production results reported closure of an approximately 2% interplant yield gap and more than 19 ingots at the highest surface rating.

The surface observations came from one furnace, and the plants differed in respects beyond arc steering. The excitation equation describes the intended current relationship, rather than a measured spatial field map.

Excitation model, installed hardware and process evidence

Engineering sketch of the twin-coil assembly showing the stationary upper coil and moving lower coil on a guided driven carriage around the mold.
FIG. 31 Engineering sketch of the twin-coil assembly: stationary upper coil, moving lower coil on a guided carriage, and its drive. AI-generated conceptual illustration.
AI-generated conceptual illustration of a cylindrical mold with fixed and moving coil assemblies, guide rods and a carriage drive.
FIG. 32 Retrofit arrangement, redrawn as an engineering sketch. A stationary coil and a coil on a guided carriage surround the mold; the carriage follows the growing product. AI-generated conceptual illustration.
AI-generated conceptual illustration of a control cabinet containing a PLC, terminals, drive electronics and DC power supplies.
FIG. 33 Engineering sketch of the control cabinet’s component groups. PLC and interface modules, terminals, carriage and coil-drive electronics, and DC power supplies are integrated in one enclosure. AI-generated conceptual illustration.
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. 34 Verification from the process record (values blurred). 1 Product height over the run, which sets the carriage position. 2 Arc rotation measured from the process camera, holding at the commanded speed.

Representative surface results

AI-generated illustrative example of a cast ingot with flaking sidewall material and a heavy irregular crown.
FIG. 35 Engineering sketch of the rough cast surface. Flaking sidewall material and a heavy crown at the rim show material removed during finishing. AI-generated conceptual illustration.
AI-generated illustrative example of an ingot with a continuous as-cast sidewall and a restrained rim.
FIG. 36 Engineering sketch of the reported improved surface condition: a continuous sidewall along the ingot’s length with reduced surface defects. AI-generated conceptual illustration.

Architecture & assurance

From one system to repeatable deployment

Herscovici connected instrumentation, computational methods and equipment control through defined interfaces. Python-based applications, PLC integration and electromechanical hardware formed systems that could be verified, maintained and deployed across facilities.

Architecture, verification and deployment footprint

Integrated hardware, software & control

Integrated hardware, software and control architecture: instrumentation and optical sensing have separate acquisition paths; Python applications integrate data and produce measurements and control requests; PLCs retain interlocks and drive actuators, with direct equipment feedback. Operations, engineering development and verification support the system.
FIG. 37 Representative architecture across deployed systems. Acquisition and interfaces vary by application. Software supplies measurements, estimates and control requests; PLCs retain interlocks and hardware control.

OPC UA connected applicable software services to PLCs; historians and MES associated results with operating records.

Engineering principles

  • Define authority and failure behavior. PLCs retained process interlocks and hardware control. Each application defined when software outputs informed operators, requested action or blocked progression.
  • Separate recognition from measurement. Neural networks located features; deterministic geometry calculated dimensions and ratings, keeping the measurement logic inspectable.
  • Verify against independent evidence. Evaluation used injected events, camera-to-laser comparisons and engineering judgments. FMEA, hazard analysis, acceptance testing and offline model validation preceded deployment; runtime monitoring and incident response supported operation.
  • Design for the environment. Installations accommodated heat, arc light, particulate, vibration, vacuum interfaces, electrical noise and cable runs of 100–300 metres.

Reuse the architecture; validate each application

Common hardware, drawings, software and procedures supported deployment across ten facilities. Site-specific work addressed mounting, thresholds, displays and process logic. Capabilities were proven at an operating site before adoption into the shared standard.

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, greenfield202610 plants
FIG. 38 Scaling by instantiation. One reference design; each site is a copy with a small site-specific slot, kept in sync by governance. Red: growth from one unit in 2016 to ten plants in 2026.

The installed infrastructure included 150 cameras, 20 thermal imagers and 35 laser systems, with more than 25 shielded, water-cooled housings behind vacuum-rated viewports.

Deployment footprint

North America1234567Europe8910
  1. 1Henderson, NV Mold inspection · flatness measurement · alignment
  2. 2Wyman-Gordon, Houston, TX Sheet inspection · bar inspection
  3. 3Huntington, WV Mold inspection · magnetic arc control
  4. 4Ravenswood, WV Mold inspection · magnetic arc control · flatness measurement · alignment
  5. 5Toronto, OH Sheet inspection · bar inspection
  6. 6Morgantown, PA Mold inspection · magnetic arc control · flatness measurement
  7. 7New Hartford, NY Mold inspection · magnetic arc control
  8. 8Swansea, UK Sheet inspection · bar inspection
  9. 9Witton, UK Mold inspection · flatness measurement
  10. 10Savoie, France Mold inspection · magnetic arc control · flatness measurement
FIG. 39 Ten plants in the U.S., UK and France and the systems deployed at each, delivered since 2020 by his R&D organization.

Experience & recognition

A background across mechanical engineering and computer science, supported by patents, publications and more than a decade of technical progression.

Patents, publications and awards

Titles are reproduced as filed. Patents are assigned to a Precision Castparts subsidiary. Herscovici served on the Liquid Metal Processing & Casting Conference committee and led annual machine-learning and computer-vision workshops at PCC.

Career, education and technical tools

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

Technical tools

  • Perception and MLPython, PyTorch, TensorFlow, OpenCV, scikit-learn
  • Inference and deploymentONNX, TensorRT, GPU servers, edge devices
  • Data at scaleDatabricks, Spark, MLflow, Azure
  • Controls and integrationAllen-Bradley PLCs, OPC-UA, Ignition HMI, AVEVA MES
  • Design and simulationSolidWorks, COMSOL, MATLAB, Simulink

Earlier production work included C/C++ and CUDA; recent development centered on Python.