Data Science and CS student at UC Berkeley, from Ukraine. I build GPU kernels, data pipelines and full-stack products, and contribute to open-source geospatial, security and data tools.

vborysenko.com · LinkedIn · Resume · volodymyr_borysenko@berkeley.edu

What I work on

• GPU and performance: CUDA C++ kernels, memory-layout tuning, benchmarks with exact parity against CPU references.
• Data at scale: country-scale crawlers and geospatial pipelines, checked for missed or duplicate records.
• Full-stack products: Next.js, TypeScript, PostgreSQL, Python services, OAuth and tested authorization.
• Modeling: probabilistic forecasting, Monte Carlo and calibration on real outcomes.

CUDA C++ kernels, compiled at runtime with NVIDIA cuda.core, that count heat-hazard days and longest runs over [scenario, location, day] arrays.

• Profiling on a Tesla T4 showed a transpose took 261 of 458 ms end to end. A shared-memory tiled kernel reading the original layout removed it: 457 ms → 208 ms (2.2×) on 187M cells.
• Every GPU result matches the NumPy reference exactly. 171 GPU tests pass on the T4.
• Baselines: NumPy, multithreaded C++17 and PyTorch. Tables are generated from each run's JSON, including where the GPU does not win (end to end, it roughly ties 10 laptop threads once data crosses PCIe).

CUDA C++ Python cuda.core NVRTC PyTorch C++17

Checks a photo of a California hospital bill against the hospital's published prices and state charity-care law, then drafts letters the patient approves. Solo; won Apify Best Actor and the Convex sponsor prize.

• Public Apify Actor (Python) that finds and streams hospital price files. Line-level price checks cover 340 of 470 California hospitals.
• Statewide audit of all 470 HCAI-listed hospitals' charity-care policies, quoting each conflict with state law.
• Eval on synthetic bills priced from real price files: 181 of 181 lines read correctly on 30 bills (not patient bills).

TypeScript Next.js Convex Python Claude API Expo

Prices every Gemini daily high-temperature contract in 49 US cities by conditioning 82 GFS and ECMWF ensemble members on that day's observed temperatures (4,000-draw Monte Carlo).

• Replayed on 978 settled markets: Brier 0.306 at 6 pm vs 0.477 for the forecast alone. All 2,345 contracts the observed floor ruled out settled NO.
• Historical replays and a paper order book, not live trading.

TypeScript Next.js Monte Carlo x402

Research and relationship tool for investors: search Gmail, Contacts and Calendar history, score companies against a thesis, and draft outreach that cites evidence.

• Google OAuth with AES-GCM encrypted tokens. Every query is user-scoped, and nothing is sent without a second confirmation.
• Python FastAPI retrieval service on PostgreSQL and pgvector (HNSW cosine search).
• CI runs TypeScript and Python suites, including PostgreSQL integration and tenant-isolation tests.

TypeScript Next.js PostgreSQL Prisma FastAPI pgvector

Also: Schmidt DSE climate pipeline (Berkeley research: xarray and Dask on downscaled CMIP6 for National Park Service maps) · Waves (heatwave risk map for Ukraine) · SwarmDAQ (task market for AI agents, WeaveHacks 2025)

57 merged pull requests in 12 external repositories (as of Oct 27, 2026; own repos excluded). The largest are below.

All the Places · 41 merged: 39 new spiders and 2 fixes. The spiders reported 239,973 locations in test crawls (raw crawl output, not deduplicated or checked as open).

• Royal Mail: the API returns only the 60 nearest places, so the spider splits the UK into quadrants until each box is covered. 136,489 places. In central London it matched a dense grid's 923 places with 78 requests instead of 390.
• Japan Post fix: post box numbers repeat across regions, so the duplicate filter dropped about 73,000 real boxes. Also fixed paging in the shared e-map crawler.
• NTT public telephones: 69,357 payphones. Converted the old Tokyo datum to WGS 84; median distance to matching OpenStreetMap phones fell from 112 m to 8 m.
• Packeta: 15,929 pickup points in 5 countries, splitting geohash cells at the 100-result cap.
• Coles fix: recursive search in dense areas, where the API caps results at 15. Inner Melbourne went from 58 to 146 stores, the exhaustive count.

OWASP Java Encoder · 5 merged

• Encode.forJson: new public API for JSON string contexts (RFC 8259), with JSP and Jakarta tags. 1,865 lines across 29 files.
• JavaScript encoders now escape DEL and C1 controls. OSGi import ranges derived from bytecode so adapters can't load against an older core. Plus XML 1.1 tags and a deprecation fix.

Other merged work

• GeoPandas #3871: explode index is now int64 on every platform, fixing 32-bit failures.
• rioxarray #940: clear error for 2D spatial coordinates and a rio.shape fix.
• NVIDIA NeMo Curator #2477: boilerplate removal now handles runs at both the top and bottom.
• Kigali Sim #867: operation types in the program validator.
• Smaller fixes and docs: GeoPandas, OWASP API Security Testing Framework, OWASP CVE Lite CLI, Dependency-Track, PySAL mgwr, PySAL segregation, Numerai.

• RISE Global Winner, 2024. Schmidt Futures and Rhodes Trust.
• ISEF King Abdulaziz and Mawhiba Special Award, 2024, for TacticSquare, a UAV combat simulator (Monte Carlo, stochastic Lanchester models, RL routing). Science News Explores coverage.
• INFOMATRIX Platinum Medal, 2024, for TacticSquare. Piloted with 10 divisions in Ukraine and raised $50,000. Code not public.
• GENIUS Olympiad Gold, Environmental Science, 2024, for FireImpactology, a wildfire air-quality and health model. Code not public.
• NASA Space Apps Challenge, 2nd in Ukraine, 2023, for FireImpactology.
• All-Ukrainian Programming Cup, 3rd place and Jane Street Award, 2025.
• Hackathons: Agentic AI SF Hackathon, 1st place (FireScout, wildfire smoke agent). Cal Hacks 13.0, Apify Best Actor and Convex prize (Recount). B@BHacks 2.0, Gemini award (Highwater).
• Research: climate-risk analysis at Berkeley's Eric and Wendy Schmidt Center for Data Science & Environment.
• Kaggle: rank 7 of 566 in Store Sales forecasting (Oct 2026, rolling) and public notebooks.