Senior ML engineer · computer vision and document AI
I build computer vision and document AI software, from model development to deployment. My work includes mobile and cloud vision, document recognition, retrieval and tools for inspecting model behaviour.
The selected catalogue includes 17 merged PRs excluding one reverted change, plus 13 open proposals. Full catalogue
- Neuralink · datarepo: Six merged fixes for query correctness, Python literals, credential refresh and packaging. #73 · #74 · #68 · #67 · #57 · #59. The documentation contribution #58 was reverted in #72.
- Unidata · netCDF4: Fixed array dimensions when scientific-data selections are empty. #1499 · Interactive shape lesson.
- pydicom · Pixel encoding: Preserved packed one-bit pixels from big-endian OW input during compression. #2380.
- Tesla · Camera Kit: Proposed fixes for Android scan accuracy, resource leaks and camera reuse. 4 open PRs · #812 #813 #817 #816
- Astropy · reproject: Proposed validation of inverse coordinates in image reprojection. 1 open PR · #630 · Playable coordinate lesson
- LSST DESC · CCL: Proposed integration-range and quadrature fixes for cosmology calculations. 1 open PR · #1310
- UNSW · Nempy: Proposed corrections to battery dispatch inputs and Linux solver compatibility. 2 open PRs · #43 #44
- pydicom · Display & modality lookup tables: Proposed overflow fixes for display and modality lookup-table indexing. 2 open PRs · #2378 #2379 · Playable display lookup
- Unitree Robotics · Teleoperation & SDK: Proposed teleoperation stop-call fix and DDS security documentation. 2 open PRs · #310 #34
- Open CLIP · CoCa example in lucidrains’ fork: Proposed a corrected model-factory call in the CoCa text-generation example. 1 open PR · #2
More merged contributions
- OpenClaw · ClawPatch: Fixes crawler access and adds website checks. 1 merged PR · #95
- Keras · Examples: Adds missing setup dependencies for a documentation example. 1 merged PR · #520
- Toloka · Python SDK: Fixes the pool analytics code example. 1 merged PR · #55
- Mobile marking detector · PyTorch Lite: Updates Android inference to PyTorch Lite and its correct model loader. 2 merged PRs · #1 #2
- CompTech · who-signed: Adds Telegram photo intake and signature storage for a verification prototype. 2 merged PRs · #2 #3
- Random Face · Colab setup: Repairs model downloads and adds a runnable Colab entry point. 1 merged PR · #1
- YParser · English documentation: Makes image-collection setup and usage available in English. 1 merged PR · #1
- Agnitra: Python SDK and CLI for model profiling, runtime telemetry and inference optimization. PyPI release · Recorded profiling example in 3D
- CV Repro Lab: Reproducible computer-vision experiments across Colab, Kaggle and GPU VMs, with benchmarks and reviewable run artifacts. Source & examples
- Artifact Redactor: Redacts sensitive text patterns from logs, Markdown and JSON; flags unsupported files for review. Source & examples
- Vehicle Lab: Open-source notebook for CAD inspection, design revisions and recorded simulation. Interactive notebook
Calorio: AI meal logging in Telegram
I built and maintain a service that helps people keep a food diary using meal photos, voice messages and text. I handle multimodal recognition, the logging workflow, deployment and technical support.
Try the bot · Engineering case study
- Riverstart · Jul 2026–present: Document extraction, retrieval and specialist review for engineering documents. Document AI
- Carb Manager: Shipped food recognition and nutrition-label extraction across cloud, iOS and Android.
- CFT: Built financial-document recognition models, assisted annotation and mobile meter-recognition workflows.
- Dermaself: Developed guided capture, pore and wrinkle segmentation, and mobile/API integration for cosmetic skin analysis. Case study
- Video search: Delivered ASR/OCR pipeline upgrades and Qdrant retrieval combining BM25 with dense embeddings. Case study
- LigninQC: Rerun two published lignin-chemistry analyses offline, with source-linked tables, code and explicit limits.
- SectionCheck: Inspect affine image transforms and polygon regions, then export explicit review decisions using synthetic images. Explore in 3D
- Drawing → catalog → interior: Follow an eight-stage recorded workflow and inspect adjustable room fit. The guide and film are available; live generation is paused.
- Engineering analysis: Professional experience with computer vision for engineering drawings and 3D geometry; private company R&D.
- ML & vision: Python · PyTorch · OpenCV · OpenMMLab · ONNX Runtime
- Systems: FastAPI · PostgreSQL · Qdrant · Docker · GCP · ClearML
- Mobile: Flutter · Dart · C++