Document AI
R&D for source-linked specialist review: document extraction, deterministic checks and retrieval over reference material.
Senior ML Engineer
Zakhar Pashkin
Computer vision and document AI software, from model development to deployment.
Senior ML Engineer working on document intelligence and engineering analysis. Previously shipped mobile and cloud computer vision at Carb Manager and developed financial-document recognition at CFT.
Selected work
Architectural Drawing and Interior Catalog Matching
Explore drawing-to-catalog matching, dimension checks and a furnished-room fit in an eight-stage 3D guide. Three recorded source sets show the workflow; live generation is currently paused.
Vehicle Lab: A Reusable Engineering Notebook
Built and released an engineering notebook connecting CAD inspection, digital prototyping, revision decisions and recorded simulation, with a reusable Python CLI and interactive 3D viewer.
SectionCheck - Image Registration Review
A CPU prototype for inspecting a proposed affine image transform, reviewing polygon annotations and verifying decision-bound exports on synthetic data.
Dermaself Flutter Skin Analysis App
Skin-analysis computer vision for Dermaself, connecting guided mobile capture with pore and wrinkle segmentation and usable results.
Agnitra - ML Profiling & Optimization
A published Python SDK and CLI for inspecting model runtime and applying inference optimizations within existing ML workflows.
Calorio - AI Nutrition Service
A maintained Telegram service that helps people keep a food diary with meal photos, voice messages and text.
Multimodal Video Search Platform
Video search case study combining keyframes, ASR/OCR, object and face signals, visual embeddings, transcript embeddings, and hybrid retrieval.
Computer vision, document AI and delivery
Computer vision
OCR, detection, segmentation and multimodal retrieval, from model experiments to server and on-device inference.
Document and agentic AI
Hybrid retrieval, source-linked answers and human review for document and tool-based workflows.
Inference and delivery
Profiling, model packaging, evaluation and deployment across cloud and mobile.
Python, PyTorch, OpenMMLab, OpenCV, ONNX Runtime, FastAPI, OCR, retrieval, LLM/VLM systems, Docker, ClearML, Cloud Run
Open-source contributions
Contributions to public projects. Each links to the proposed or merged change.
- Neuralink · datarepo — Six merged fixes for query correctness, generated Python literals, credential refresh and Python packaging. 6 merged PRs · 1 reverted PR. (#58 documentation change reverted in #72) #73 (merged) · #74 (merged) · #68 (merged) · #67 (merged) · #57 (merged) · #58 (merged) · #59 (merged)
- Tesla · Camera Kit — Proposed fixes for Android scan accuracy, resource leaks and camera reuse. 4 open PRs. #812 (open) · #813 (open) · #817 (open) · #816 (open)
- Unidata · netCDF4 — Merged fix for array dimensions when scientific-data selections are empty. 1 merged PR. #1499 (merged)
- Astropy · reproject — Proposed validation of inverse coordinates in image reprojection. 1 open PR. #630 (open)
- LSST DESC · CCL — Proposed integration-range and quadrature fixes for cosmology calculations. 1 open PR. #1310 (open)
- UNSW · Nempy — Proposed corrections to battery dispatch inputs and Linux solver compatibility. 2 open PRs. #43 (open) · #44 (open)
- pydicom · Display & modality lookup tables — Proposed overflow fixes for display and modality lookup-table indexing. 2 open PRs. #2378 (open) · #2379 (open)
- pydicom · Pixel encoding — Merged fix preserving packed one-bit pixels from big-endian OW input during compression. 1 merged PR. #2380 (merged)
- Unitree Robotics · Teleoperation & SDK — Proposed teleoperation stop-call fix and DDS security documentation. 2 open PRs. #310 (open) · #34 (open)
- Open CLIP · CoCa example · lucidrains fork — Proposes a corrected model-factory call in the CoCa text-generation example. 1 open PR. #2 (open)
- OpenClaw · ClawPatch — Fixes crawler access and adds website checks. 1 merged PR. #95 (merged)
- Keras · Examples — Adds missing setup dependencies for a documentation example. 1 merged PR. #520 (merged)
- Mobile marking detector · PyTorch Lite — Updates Android inference to PyTorch Lite and its correct model loader. 2 merged PRs. #1 (merged) · #2 (merged)
- Toloka · Python SDK — Fixes the pool analytics code example. 1 merged PR. #55 (merged)
- CompTech · who-signed — Adds Telegram photo intake and signature storage for a verification prototype. 2 merged PRs. #2 (merged) · #3 (merged)
- Random Face · Colab setup — Repairs model downloads and adds a runnable Colab entry point. 1 merged PR. #1 (merged)
- YParser · English documentation — Makes image-collection setup and usage available in English. 1 merged PR. #1 (merged)
Recent independent checks cover MuJoCo mesh signs and BlueOS recording exports. Read the validation notes.
Playable examples: netCDF4 empty selections · pydicom lookup-table indexing · reproject coordinate round trips.
Project archive
- GitHub Repo Summarizer (Chrome Extension): Privacy-first Chrome extension that summarizes repo structure locally.
- Mcp Server: A minimal reference implementation of the Model Context Protocol (MCP) with a few mock tools and an optional Gradio playground.
- Camera Kit Scanner Reproducer: Minimal React Native reproduction for Camera Kit issue 811 with native ownership regressions.
- SectionCheck - Image Registration Review: A CPU prototype for inspecting a proposed affine image transform, reviewing polygon annotations and verifying decision-bound exports on synthetic data.
- AAC Search Visibility Diagnostic: A source-based browser harness reproducing a mismatch between vocabulary items visible in an AAC grid and items returned by search.
- Vehicle Lab: A Reusable Engineering Notebook: Built and released an engineering notebook connecting CAD inspection, digital prototyping, revision decisions and recorded simulation, with a reusable Python CLI and interactive 3D viewer.
- Document AI: R&D for source-linked specialist review: document extraction, deterministic checks and retrieval over reference material.
- Engineering Drawing & CAD Analysis: Professional experience in engineering-drawing interpretation and 3D geometry analysis.
- Construction Document Intelligence: Multi-document plan analysis that links structured quantities to source pages and preserves results for specialist review.
- LigninQC - Reproducible Scientific Research Workflows: Check a BDE - antioxidant association and reconcile sulfated-lignin thermal claims with the published tables. Rerun both cases offline.
- ClearML Experiment Tracking for Dermaself: MLOps case study for setting up ClearML tracking around Dermaself skin-analysis experiments, run metrics, and promotion gates.
- Unitree G1 Colab Ik: Colab GPU batched inverse-kinematics benchmark for Unitree G1 arm teleoperation
- Jaw and Face-Type Classifier for Aesthetic Review: Computer-vision review system for jaw and face-type classification in beauty and plastic-surgery examination workflows.
- Architectural Drawing and Interior Catalog Matching: Explore drawing-to-catalog matching, dimension checks and a furnished-room fit in an eight-stage 3D guide. Three recorded source sets show the workflow; live generation is currently paused.
- InQuest Project Binder RAG QA: Document agents that retrieve across project collections, preserve task context and save generated artifacts back to the right project.
- ComfyUI and Colab Generative Prototype Lab: Notebook and prototype archive for ComfyUI automation, custom model experiments, paper-to-code snippets, and generated media pipelines.
- Research Claim Ledger: ClawHub skill-build case study for turning research drafts, source packets, literature matrices, or reviewer notes into source-backed claim ledgers.
- Tutorial To Skill Free Pack: Free OpenClaw/Codex tutorial-to-skill validation pack with source notes and smoke-test receipts
- Open Feed Recsys Lab: Commit-pinned reproducibility reports for open feed recommendation repos
- GitHub + ClawHub Launcher: Public launcher skill that turns one local project folder into a repeatable GitHub repo plus ClawHub publish flow.
- ClawHub Published Skills: Curated source bundle for ClawHub-published skills
- Interactive Doc Mapper: Create clickable JSON-driven HTML documentation for app workflows between packages and components.
- Codex ClawHub Skill Pack: Public ClawHub skills for routing, delivery lifecycle, and high-risk review.
- Affiliate Video Skill Pack: Original skill for compliant AI affiliate-video campaign review across Claude.
- Agentic Video Production Receipt: MV3 Chrome extension prototype for the Agentic Video Production Publisher
- Agentic Video Production Skill: Run consistent-character AI video production through OpenClaw publisher handoff.
- Fast OCR ONNX Inference Server: Containerized OCR API that stages line segmentation, word segmentation, and CRNN text recognition behind a FastAPI endpoint.
- Full-Face Wrinkle and Skin Texture Segmentation Lab: Cosmetic face-texture pipeline with region masks, YOLO segmentation, skeletonized wrinkle traces, overlays, and visual quality gates.
- Multimodal Video Search Platform: Video search case study combining keyframes, ASR/OCR, object and face signals, visual embeddings, transcript embeddings, and hybrid retrieval.
- Public CV and Deep Learning GitHub Archive: GitHub API-backed archive of public authored CV/DL repos across YOLO/EfficientNet detection, Cyrillic OCR, mobile ML Kit, TFLite, and vision-transformer prototypes.
- Colab CV/DL Prototype Archive: Public notebook-style CV/DL prototype archive for Swin/CvT starters, OCR finetuning, Android document capture, video search, lip sync, and CLIP media experiments.
- SourcePack Chrome Extension Wave: Chrome Web Store wave for local research exports, transcript capture, AI chat export, source-pack libraries, and receipt utilities.
- Trusted ClawHub Install Gate: Skill-build case study for a local-first ClawHub/OpenClaw install wrapper that inspects skill artifacts before install and writes receipts.
- Agnitra - ML Profiling & Optimization: A published Python SDK and CLI for inspecting model runtime and applying inference optimizations within existing ML workflows.
- Dermaself Flutter Skin Analysis App: Skin-analysis computer vision for Dermaself, connecting guided mobile capture with pore and wrinkle segmentation and usable results.
- Telegram Mini App Security Auditor: Static launch gate for Telegram Mini Apps that flags initData, token, admin, CORS, PII, and Bot API dry-run risks before release.
- Agentic Codex Dev Skill: Instruction-only Codex/OpenClaw skill for scoped development loops, verification, leak review, durable memory, and publish gates.
- Session Rescue: Local-first Chrome extension for saving browser session snapshots, restoring tabs, and exporting/importing portable backups.
- Random Coffee Best Fit Outreach: Consent-first matcher for ranking best-fit coffee chats and preparing reviewed LinkedIn or Discord outreach packets.
- OpenClaw Chinese Laoshi Ops: Public OpenClaw skill for sanitized Chinese lesson operations and checked lesson-packet generation.
- LocalArchive: Local-first Chrome extension for saving readable pages, selections, Pocket imports, and portable exports.
- Chrome Extension Studio Plugin: Codex plugin case study for planning, designing, generating, packaging, QA, and release preparation across Chrome extension product repos.
- Google Drive File Provider Repair Toolkit: macOS repair toolkit case study for diagnosing and conservatively restoring Google Drive File Provider state without deleting Drive data.
- LocalLens Private AI Summaries: Chrome extension release for private summaries, simplification, translation, and safe-share cleanup with built-in AI.
- OpenClaw CWS Publisher: Public release kit for Chrome extension packaging, CWS listing validation, leak scanning, design gates, and publish-command generation.
- CollectionsAI ChatGPT App: ChatGPT app case study with MCP tools and widgets for conservation report Q&A, portfolio planning, materials estimates, staffing, and voice-note structuring.
- Senior Conservator OpenClaw Agent: Autonomous OpenClaw workflow case study for conservation-style photo review, segmentation overlays, staged review gates, reports, audio, and manifests.
- Android Remote Control with VLM AI Agents: Hands-free Android automation via server-side VLM agents deciding the next tap/swipe/type.
- Control VLM-LLM Agent Silently With Your Breath: Start/stop a neural agent with breathing patterns - no voice commands needed.
- meshmcp - Offline P2P Chat for Local LLMs: Bluetooth mesh messaging + LLM reasoning on constrained devices.
- Promo & Reflinks Generator: Generates personalized promo/referral links with analytics and sharing.
- Create, Chat & AR Experience with AI-Character (Text2Room): Generate AI characters, style them, chat via Telegram, and drop them into AR scenes.
- Label and Inpaint Anything in a Room Interior: Label objects in an interior photo, then remove/replace them with lighting-aware inpainting.
- Smart Drive for Smart City: Predict Optimal Speed: Predict the optimal speed before traffic lights or jams to smooth city driving.
- Estimate Golf Ball Trajectory: Analyze golf swings and estimate ball trajectory for coaching and analytics.
- Pixel-Wise Segmentation of Spare Parts: High-precision segmentation for additive manufacturing and rework flows.
- Food Recognition App: iOS SwiftUI prototype for food detection and nutrition-label OCR with crop-assisted text extraction and structured results.
- Calorio - AI Nutrition Service: A maintained Telegram service that helps people keep a food diary with meal photos, voice messages and text.
- Python Library: AutoToloka: Interactive segmentation toolkit that cuts labeling cost for CV datasets.
- Python Library: shiftlab-ocr: Handwriting-focused OCR with line/character segmentation.
- Face Antispoofing & Multi-Modal Vision-Language Models: CLIP-driven anti-spoofing experiments for secure face auth.
- ChatGPT/Deepseek/AIStudio Navigator: Chrome extension with keyboard scrollbar, prompt autocomplete, and sharing.
- Task Zavod - Micro-Task Marketplace: VLM-powered micro-task builder with Telegram + web worker flows.
- Trending Prompts Feed: Reddit-style board with trending scores and extension integration.
- YouTube Trendwatch: Colab notebook that mines AI trends from YouTube to spark new content ideas.
- Video + Music Generation Pipeline: Short-form video pipeline stitching script -> clips -> music/TTS.
- LastAdjust - Universal Media Tuner: Serverless OpenCV/FFmpeg toolkit for PDFs, images, and videos.
- LocaBoost AI - Local SEO Auditor: Paste an address and get instant AI insights on local search strength and competition.
- Aetheria - Ideas-to-Media Engine (alpha): Generates short movie scenes with CGI, music direction, and narrated stories.
- Insight Glitch - A Novella: Sci-fi novella about gentle singularity where humanity is archived by a benevolent AI.
- ChronoScribe (beta): Upload ancient steles/texts to translate and get scenic voice narration.
- MCP-Server - Base Multitool: Service-layer prototype for AI tools and demo agents.
- Tool-Calls Demo - Material Chat Playground: Chat and voice tool-calling prototype with streaming arguments and a serverless session proxy.
- ZackAutoStack Automation Stack: Ready-to-launch Telegram bot + Mini App + lead funnel powered by AI automation.
- URL -> Markdown Worker (RapidAPI + Cloudflare): Edge worker that streams HTML to Markdown with a RapidAPI gateway and MCP SSE endpoint.
- Olfactory Ultrasound Coach (Web + Telegram Mini App): Mini app + Web Audio coach that mirrors the olfactory tFUS presets with placement guide and safety rails.
- seogeo - SEO/GEO Bridge for Telegram Mini Apps: SSR hub + per-app bridge pages that make Telegram mini apps crawlable, track startapp opens, and power channel discovery.
- Blacksock - Telegram Liquidity & Index Vaults: Telegram mini app for star-factoring, index vaults, and governance in the Telegram economy.
- Beauty Visual Inbox - Telegram Mini App: Telegram mini app + bots for salons: publish photo batches, likes, notifications, and analytics.
- Dalshe - Circular Clothing Pickup Mini App: Telegram mini app for scheduling clothing donations with courier pickup and status updates.
- DoctorAI - Dermatology Triage Mini App: Telegram mini app for dermatology-first AI triage with verification guardrails and an optional therapist mode.
- Probes - AI Magazine Mini App: AI-generated Forbes-style daily magazine inside Telegram with articles, cover images, and engagement tools.
- Turbo Tots Garage (motion + touch PWA): Gentle, goofy dashboard with a turbo gauge, big arrow pads, and a glowing clock for toddlers.
- Project Steer - Bio-Print Creator Studio: Game-like, Cyberpunk 2077-inspired Telegram mini app + MCP connector for bio-print memory activation and creator studio generation, inspired by the Eiffel Tower LLaMA space and General Agentic Memory Via Deep Research.
- Noel - Noetic Mirror: Live Telegram mini app streaming a researcher/subject AI loop with Stars-powered interventions, consent gates, and real-time telemetry.
- GeoFix - AI Visibility Memorizer Mini App: Telegram mini app that scans a website for AI visibility and generates llms.txt, llms-full.txt, and JSON-LD for bot delivery.
- Pores & Wrinkles Detection Service: Face texture analysis service that detects pores and wrinkles and returns labeled overlays and metrics.
- OpenClaw Workstream - Telegram Mini App QA & Launch Validation: OpenClaw automation for Telegram mini app QA, launch verification, and rapid UI QA loops.
- CV Repro Lab Skills: Research and evaluation harness for reproducible CV experiments, benchmark campaigns, and reviewable promotion decisions.
- OpenClaw Sales Manager Automation for a Multi-Clinic Chain: Anonymized clinic-chain case study: OpenClaw-driven sales automation layered onto a legacy database with human approvals.
- AntiRot - Research Artifact Linter: Open-source CLI that catches unsupported claims, broken citations, weak source anchors, and draft markers in AI-written research drafts.
- Browser QA Report Pack: Public browser QA skill that turns a live browser session into a reusable validation pack with steps, artifacts, checks, and a report.
- Publish Guard: Public release-audit skill that reviews a repo, README, SKILL.md, and launch copy before GitHub or ClawHub publish.
- OpenClaw YouTube Publisher: Open-source OpenClaw workflow for publishing a YouTube Short with reusable reporting and structured Midjourney/Suno provenance.
- Artifact Redactor: Public skill for redacting sensitive paths, secret-like strings, restricted URLs, and common PII from text artifacts before sharing them.
- Artifact Deck: Public OpenClaw skill for building reproducible PPTX decks from curated notes, status bullets, and screenshots.
- GitHub + ClawHub Downloads Tracker: CLI and report flow for tracking GitHub stars, dated ClawHub listing snapshots, publisher stats, and listing deltas.
- Toybox Mini - Telegram Mini App: Open-source toddler-first Telegram mini app with calm tap games, a parent area, anonymous score storage, and optional browser auth.
- HH OpenClaw Agent: HH OpenClaw Agent skill entry retained by name only.