Senior AI / ML Engineer — Production ML Systems · Computer Vision · Edge Inference & ML Evaluation
📧 piotr.gryko@gmail.com
🔗 LinkedIn • GitHub • GitLab
📍 Warsaw, Poland · open to relocation
📄 Download CV (PDF)
Summary
Physicist-turned-engineer (MSci Physics UCL, PhD Imperial College): 12+ years in engineering, 5+ in applied ML, 3+ shipping deep-learning systems that survive production — GPU inference fleets, evaluation infrastructure, data engines, and edge computer vision under hard latency budgets. Contractor via my consultancy DevaLogic: Hypodossier is my primary engagement, alongside a counter-UAS programme for a defense-tech client fed by my own drone dataset engine. I own systems end-to-end — model serving, distributed pipelines, security hardening, and the measurement discipline (confidence intervals, cost evals, reproducible runs) that separates shipped ML from demos. Conference speaker (EuroPython, PyCon Lithuania, Data Science Summit).
Core Skills
ML engineering: PyTorch, transformers, VLMs (Gemini, Qwen2-VL, SigLIP 2), YOLO/detection & tracking, diffusion models, embeddings, LayoutLM, scikit-learn; rigorous model evaluation — golden datasets, FROC/Pd/FAR curves, calibration, bootstrap and exact confidence intervals, cost-per-inference analysis; experiment tracking & feature stores (MLflow, Feast)
Systems & performance: GPU inference optimization (CUDA, TensorRT), NVIDIA Jetson edge pipelines, C/C++, HPC signal processing, async/distributed architectures (RabbitMQ incl. mTLS, Hatchet durable execution), memory-leak and throughput forensics
Platform & infra: Python, PostgreSQL/pgvector, Docker, Kubernetes/ArgoCD GitOps, Pulumi IaC, supply-chain security (Trivy, OpenVEX, SBOM), AWS/GCP; Django/FastAPI/React product layers
Experience
R&D Software Engineer (ML Systems) | Hypodossier — contract via DevaLogic | May 2023 – Present
AI document-intelligence platform for the Swiss mortgage-banking industry. ~7,000 commits across 34 services over 3 years; largest single contributor to the core product monorepo.
- Built and operated a fleet of GPU inference microservices over RabbitMQ: LayoutLM extraction, EfficientNet page classification, Qwen2-VL visual document embeddings, and page-stream segmentation running at <6.25 ms/page — with reproducible CUDA/PyTorch/TensorRT base images for the whole ML fleet.
- Sole author of the internal ML labeling and evaluation platform: golden-dataset lifecycle with versioned snapshot releases, per-class F1 and calibration dashboards, KNN mislabel detection over pgvector embeddings — then used it to drive a shadow-mode → primary rollout of a new production page classifier.
- Delivered measured performance wins: redesigned a vision-language embedding pipeline for a 25% speedup by decoupling preprocessing from GPU inference; diagnosed production memory leaks and RabbitMQ channel-death stalls with A/B soak-test harnesses; added cost-control guards preventing silent commercial-OCR billing on cache misses.
- Led the org-wide container-security programme: standardized Trivy + pip-audit + OpenVEX CI gates and runtime hardening across 15 image-building services, driving scans to zero active findings; rolled out fleet-wide RabbitMQ mTLS with fail-closed design.
- Built supporting services end-to-end: a national building-registry search service (3.3M buildings, Meilisearch), multi-engine antivirus scanning, PDF/A conversion, and bank-tenant integration APIs on the core Django/React product.
Founder & Principal ML Engineer | DevaLogic (own consultancy) | Jan 2023 – Present
My consultancy — the contracting vehicle for the Hypodossier engagement above, plus the following client and product work:
Counter-UAS Computer Vision — Software & AI Lead (client engagement) · Jun 2026 – Present
Ongoing R&D for a defense-tech client: real-time detection of very small (sub-pixel-to-few-pixel) aerial targets in compressed video on Jetson-class edge hardware. The v1 detector is in development; the evaluation harness and methodology were built first.
- Architected a classical-CV detection pipeline (video-only stabilization, CFAR-style adaptive thresholding, velocity-hypothesis track-before-detect) under a hard end-to-end latency budget; authored the technical research plan from the small-target detection literature.
- Built the evaluation harness before the detector: contract-driven evaluator (135 tests, golden fixtures) producing probability-of-detection, FROC, and false-alarm-rate metrics with proper confidence intervals, plus hashed-config run logs making every published number reconstructible.
Drone Detection Dataset Engine — own product, sole engineer · Jul 2025 – Present
Designed, built, and operate solo a production ML data engine converting publicly posted video into annotated computer-vision training datasets — running unattended in production.
- Automated pipeline: multi-source ingestion (220k+ videos, 1.5 TB) → vision-language-model batch triage (118k+ analyses) with compilation detection and frame-accurate cutting → embedding-based keyframe selection → segmentation-model-assisted annotation in CVAT → YOLO/COCO/VOC export.
- Built a model-migration evaluation harness (400-clip labelled fixture, bootstrap confidence intervals, measured cost) that disproved a “cost-neutral” migration assumption by 2.6× and caught an accuracy collapse in AI-generated-content detection (0.875 → 0.375) before it contaminated training data.
- Production operations as a one-person team: durable workflow execution across CPU/GPU/ingest workers, infrastructure-as-code deployment with zero public ports, tiered backup verification with weekly restore drills, 2,000+ tests in CI.
- Shipped the commercial layer: per-client rate cards over seven billing metrics, freeze-at-ship invoice snapshots, source-stripped client dataset bundles.
Earlier consulting projects · 2023 – 2025
- Document anonymization with diffusion models (VAE/UNet inpainting per the DiffUTE paper) — presented at EuroPython 2025 and PyCon Lithuania 2025.
- Full-stack RAG system with self-hosted LLMs (async Django REST API with streaming responses and WebSockets, React/TypeScript frontend, vector DB).
- Property-suitability ranking model for a UK energy-retrofit client (2024, two-person team): EPC open data joined to survey outcomes for 13,000 properties, 173 engineered features incl. NLP over free-text building descriptions; Feast feature store + MLflow experiment tracking, SMOTE, Optuna-tuned XGBoost — F1 0.78.
Senior Software Engineer | Qogita | Mar 2022 – May 2023
Helped scale the engineering behind a B2B marketplace doing ~$20M/month in turnover: payment-system migration (Square → Wise) with zero outages, two-way Snowflake–Django data sync, high-throughput ingestion pipelines.
R&D Software Engineer | Unipart Digital | Oct 2015 – Jan 2022
Multidisciplinary R&D for a large logistics group: classical ML (demand forecasting, telemetry anomaly detection; scikit-learn, pandas), web systems (Django/React), devops (Docker, Ansible, Elastic), embedded C++ (Arduino, ESP32); led teams of 3–4 engineers.
R&D Software Engineer | ION Geophysical | Jun 2013 – Oct 2015
Signal-processing tools for seismic imaging on high-performance computing clusters (C++); integrated Python/NumPy into the C++ processing system.
Education
PhD, Self-Assembling Biomaterials / Nanotechnology · Imperial College London · 2007–2012
X-ray scattering physicist: in-situ synchrotron SAXS on polypeptide-bridged gold-nanoparticle assemblies for bio-sensing, fitting form-factor and structure-factor models (sticky-hard-sphere, para-crystalline lattice, fractal-dimension analysis) to link nanoscale structure to optical response.
MSci Physics, First Class Honours · University College London · 2003–2007
First-author paper on coherent X-ray diffraction imaging of gold nanocrystals at the Advanced Photon Source: wrote C/Python simulations that explained an unexpected doubled diffraction pattern as standing-wave illumination, then recovered real-space images by iterative phase retrieval.
Publications
- Aili, D., Gryko, P., Sepulveda, B., Dick, J. A. G., Kirby, N., Heenan, R., Baltzer, L., Liedberg, B., Ryan, M. P., Stevens, M. M. — “Polypeptide Folding-Mediated Tuning of the Optical and Structural Properties of Gold Nanoparticle Assemblies”, Nano Letters 11(12), 5564–5573 (2011). doi:10.1021/nl203559s
- Gryko, P., Liang, M., Harder, R., Robinson, I. K. — “Observation of interference effects in coherent diffraction of nanocrystals under X-ray standing-wave illumination”, Journal of Synchrotron Radiation 14, 471–476 (2007). doi:10.1107/S0909049507040228
Talks
“Anonymization of Sensitive Information in Financial Documents” — EuroPython 2025 🎥 Watch, PyCon Lithuania 2025, Data Science Summit: ML Edition 2025
“Building and Scaling an AI Startup with Async Django” — PyCon Lithuania 2024 🎥 Watch
“Scaling, Refactoring and Fixing a Django MVP for Production” — PyCon Lithuania 2024 🎥 Watch
Community
Founder & Organizer — AI Code & Coffee Warsaw tech mentoring meetup
Open source — GitHub · GitLab
Open to senior ML / AI engineering roles (computer vision, ML platforms, applied ML) and to consulting and speaking engagements.