Rukesh DuwalCOMPUTER VISION ENGINEER
Open to CV roles · Melbourne
STATION_00 · ONLINErukesh duwal / computer vision engineer / melbourne, au

I teach cameras to catch what people miss.

Computer vision engineer with three years shipping inspection systems onto real factory lines. I was a founding engineer at Crimson Tech in Nepal and part of the vision team behind 52 deployed applications in pharma, FMCG, food, automotive and electronics. I'm now finishing a Master of AI at La Trobe University in Melbourne.

See the deployments
RECSTATION 01 — CAP · CODE · FILLCAM-A 2448×2048 · GigE · 120 fpsyolov8n-int8.engine● LIVE SIM
Inspected0
Pass0
Reject0
Yield—
Rate / min0
Latency—

    ↑ A browser re-creation of a cap / code / fill station, the kind of inline check I built and deployed. It's simulated, not customer footage.

    52vision & IoT applications deployed by the teamCrimson Tech catalogue
    72→100%weld defect detection vs manual checksAutomotive EOL
    0customer returns in the first 8 monthsAutomotive EOL
    400 PPMbatch & expiry codes verified inlineFMCG OCR
    300 PPMblister pockets, 100% coveragePharma blister
    <50 msper-frame defect inferenceCT Garud engine
    01 / model

    How the model sees a frame.

    Every inspection runs the same chain: pixels in, features out, then boxes and a verdict. The feature maps on this rig are computed live in your browser with real 3×3 convolution kernels and 2×2 max-pooling. Move your cursor over it to turn it in 3D.

    drag / hover to rotate
    yaw −34° · pitch −10°live conv · 3 → 6 → 8 → 10 → 12 ch
    1. INPUT · 128×96A photo-eye triggers the camera. Lighting and exposure are locked at install so the model sees the same world every shift.
    2. CONV1 · edgesSmall kernels respond to edges and gradients: cap rims, label borders, weld boundaries.
    3. CONV2 · texturesPooled maps combine edges into corners and textures at half resolution.
    4. CONV3 · partsDeeper channels fire on object parts. This is where small-defect recall is won or lost, and where SAHI tiling helped on tiny parts.
    5. CONV4 · semanticsCoarse, many-channel maps that can tell a bottle from a carton.
    6. HEAD · boxes + verdictThe head regresses boxes and classes, NMS removes duplicates, and line rules turn the result into OK / NG for the PLC.
    02 / deployments

    Five deployments, start to finish.

    These are the published Crimson Tech case studies from my time on the vision team. Each one has a live camera feed, recreated in code, showing what the system checks for.

    VisionAutomotiveSegmentation

    End-of-line weld & assembly inspection

    An automotive Tier-1 supplier relied on operators to spot weld defects at the end of the line. We put the CT Garud engine on that station and wired its OK / NG output straight into their PLC.

    72→100%weld defect detection
    0returns in 8 months
    1 arvoclient-side retrain
    Problem
    Manual checks caught about 72% of weld defects. The rest came back as customer returns.
    Build
    Weld segmentation plus assembly checks on the EOL station, with each segment graded OK / loose / burnt and a PLC handshake for rejects.
    Result
    Every weld is inspected, there were zero returns in the first eight months, and the client's own team retrains new models in one afternoon.
    SegmentationCT Garud<50 msPLC / ModbusGigE
    VisionFMCGOCR / OCV

    High-speed batch code & expiry OCR

    Batch codes and expiry dates on an FMCG line were spot-checked by hand. We moved it to 100% inline verification against master data, with automatic rejection.

    400 PPMverification speed
    100%codes checked
    Inlinerejection
    Problem
    Misprinted and unreadable codes could reach dispatch, which is a compliance and recall risk.
    Build
    Retrainable OCR for inkjet, laser and emboss prints, matched to master data, plus date-logic rules (expiry after manufacture, within range).
    Result
    Every code is verified at full line speed. Smudged prints and transposed or impossible dates are rejected inline.
    CT OCRPaddleOCRDate logicMaster-data matchReject gate
    VisionPharmaceuticalsClassification

    100% blister pack inspection

    A pharmaceutical manufacturer inspected blisters by sampling. CT Mikha now checks every pocket on every card and writes the GMP record automatically.

    300 PPMinspection speed
    100%blister coverage
    3 linesstandardised
    Problem
    Sampling left a risk of empty, broken or discoloured tablets reaching the market, and GMP records were kept by hand.
    Build
    Pocket-level inspection of every blister at line speed, with a built-in GMP-compliant audit trail.
    Result
    Sampling is gone, there have been zero sampling failures since go-live, and the setup is now standard on three lines.
    CT MikhaPixel-level gradingGMP audit trailLine scan
    IoTFood & BeverageGorkha Brewery

    Live production-line tracking

    Gorkha Brewery logged output and downtime on paper. We connected line sensors and PLCs into one live dashboard covering the whole facility.

    Real-timeline visibility
    1dashboard, whole facility
    Week 1first bottleneck found
    Problem
    Stoppages only became visible after the shift, too late to act on.
    Build
    Live counts, station status and throughput streamed from the line, with each downtime event logged along with its reason.
    Result
    Tally sheets are gone, shift reports are automatic, and the real bottleneck showed up in the first week.
    Line sensorsPLCEvent loggingDashboards
    Safety AIFood & BeverageGorkha Brewery

    Digital EHS: PPE & no-go zones

    Safety compliance at the brewery was checked by walking the floor and recorded on paper. AI on CCTV now watches PPE and restricted zones in real time.

    Real-timePPE monitoring
    Instantzone alerts
    DigitalEHS records
    Problem
    No real-time view of unsafe conditions on the floor and no searchable incident history.
    Build
    Person and PPE detection (helmet, vest) on CCTV, polygon no-go zones, and IoT events feeding a digital safety log.
    Result
    PPE is checked at entry and on the line, zone breaches raise instant alerts, and paper logs are replaced by searchable records.
    Person detectionPPE classesZone polygonsRTSP / CCTV
    03 / catalogue

    All 52 applications.

    The full set of systems Crimson Tech has deployed on working lines, across eight industries. Filter by industry or search for a technique such as OCR, counting, weld or PCB.

    04 / role

    From first pilot to a 52-app catalogue.

    I joined Crimson Tech as a founding employee and stayed through its growth. I worked inside a small vision team with our team lead and senior computer vision engineer, from site visits through to the PLC signal.

    Timeline

    • Nov 2025 — Jun 2027

      Master of Artificial Intelligence

      La Trobe University · Melbourne

      Studying the theory behind what I shipped: deep learning, vision transformers, MLOps, Azure and Databricks.

    • Aug 2022 — Oct 2025

      Machine Learning / Computer Vision Engineer

      Crimson Tech Pvt. Ltd. · Lalitpur, Nepal · founding employee

      Built and deployed inline inspection for FMCG, pharma and industrial clients including Dabur, Unilever, Gorkha Brewery, Surya Nepal and TATA. Work covered detection, segmentation, OCR, counting and measurement on edge hardware under line-speed limits.

    • 2021

      Bachelor of Information Technology

      Purbanchal University · Nepal · GPA 3.5

    The loop I work in

    1. ScopeSample parts, defect list, and an honest answer to "is this detectable?"
    2. CaptureCamera, lens, lighting, triggers, calibration (px/mm)
    3. LabelLabel Studio, CVAT, Roboflow, SAM-assisted masks
    4. TrainYOLOv5/v8/v11, U-Net, DeepLabV3, PaddleOCR
    5. OptimiseINT8 quantisation, pruning, TensorRT
    6. DeployJetson Nano, Raspberry Pi, on-prem GPU, Docker
    7. IntegrateOK / NG to PLC, Modbus, FastAPI services
    8. IterateField failures become new data, then a retrain, every week
    Preprint · TechRxiv · Jun 2025

    Improved FMCG Quality Inspection using YOLOv5 and SAHI

    Slicing-aided inference to lift small-defect recall on high-resolution line images.

    Internal tool

    SAM-assisted annotation GUI

    YOLO + Segment Anything for fast instance masks, with augmentation and training-ready export. It cut labelling time by more than 60%.

    Side project

    Offline RAG assistant

    LangChain, Ollama (LLaMA) and FAISS running fully local, as a practical test of where foundation models actually add value.

    05 / stack

    Tools I've used on real lines.

    Vision models

    YOLOv5 / v8 / v11U-NetDeepLabV3SAMPaddleOCRGANs (augmentation)OpenCVSAHI

    Languages & frameworks

    PythonPyTorchTensorFlowNumPyC / C++Java

    Edge & deployment

    Jetson NanoRaspberry PiTensorRTINT8 quantisationPruningDockerFastAPIStreamlit

    Data & MLOps

    Label StudioCVATRoboflowMLflowDatabricksGit

    Industrial I/O

    GigE camerasPhoto-eye triggersPLC / ModbusRTSP CCTVReject actuators

    Cloud (learning)

    Azure MLDatabricksAWS EC2 / S3
    OPEN TO WORKgraduate / junior · melbourne or remote

    Hiring for computer vision in Australia?

    I'm looking for Computer Vision, ML or AI Engineer roles. I'm happy to walk through any deployment on this page in detail: the data, the failure modes, and what we changed after go-live.

    Emailiamrukeshduwal@gmail.com
    Phone+61 470 535 768
    LinkedInlinkedin.com/in/rukesh-duwal-544a0a193
    GitHubgithub.com/rukeshduwal