Computational scientist turned software engineer, with 8+ years across scientific computing, distributed data systems, and applied AI. Most recently I stood up the generative AI capability at a biotech company as the sole engineer on the work.

Santa Fe, NM · Open to work · U.S. citizen

Gate Bioscience — Informatics Developer II, 2024–2026. Sole engineer on most projects, embedded with research and operations staff and owning problems from framing through production.

• Built the company's generative AI capability from nothing: LLMs connected to internal databases and analysis tools over the Model Context Protocol, with the tool interfaces, structured outputs, and grounding needed for non-technical staff to rely on the answers.
• Defined the evaluation loops and acceptance criteria for AI behavior — tool invocation correctness and response quality against fixed tasks — which caught regressions before users hit them.
• Architected the production PostgreSQL schema and built Python and R pipelines on AWS, Docker, and Nextflow to ingest, validate, and integrate large datasets from external and vendor systems.

OpenEye / Cadence Molecular Sciences — Technical Developer II, 2020–2024, Santa Fe. Cloud-native scientific data platform for industrial R&D. Promoted twice.

• Led technical direction of a major product, translating requirements from application scientists, product, and customers into shipped capability.
• Architected a configurable distributed search system over billion-scale molecular databases in Go, Python, and C++ — data representation, indexing strategy, and parallelized query execution under latency constraints.
• Set development standards, mentored developers and scientists, and managed an intern across a matrixed organization.

Earlier — meshing for cloud CFD at SimScale (Munich); C++ for U.S. defense and satellite analysis at Stellar Science; subsurface flow simulation on HPC clusters at Los Alamos National Laboratory.

Education — M.S. Computational Science & Engineering, TU München · B.S. Chemical Engineering, Yale

climatelens — pick any point on Earth and see what has already changed there, computed from ERA5 reanalysis against a 1951–1980 baseline rather than recalled by a model. Includes a local El Niño composite: NOAA's index says which winters since 1950 were El Niño, ERA5 says what those winters were like at that point, detrended so the warming signal isn't double-counted, with hit rates and a permutation test on every result.

chilecule — agentic drug discovery on entirely open-source tooling: RDKit, AutoDock Vina, fpocket, and ChEMBL exposed as MCP tool servers and driven by Claude agents. Includes a retrospective validation harness that runs the agents against programs whose outcomes are already known, so the output can be scored rather than assumed.

otowi — traffic simulation for northern New Mexico: the Los Alamos commute over Otowi Bridge, and getting across Santa Fe. Most traffic models built outside transport agencies are an OpenStreetMap network plus randomly generated trips — they animate beautifully and mean nothing, because the demand is invented. Here the origin–destination flows come from census LEHD LODES data. Early, and labelled as such: nothing is calibrated yet, so no number it produces should be believed until it is.

xcpredict — predicts FIS Cross-Country World Cup results from athletes' prior races, driven by the published start list for the specific race. A season-long ranking will happily tell you the favourite wins Sunday's 20 km; it does not know they skipped the trip. Elo ratings feed a Monte-Carlo simulation, and the field comes from FIS itself, so a prediction never contains someone who is not starting.

polarswim — Polar computes per-length swim data in its cloud, shows it in the web app, and omits it from every file export. This pulls it out of the private API into local SQLite. 7,600 lengths and counting.

jobscout — an agentic job finder that reads the applications you've already written, searches employers' own ATS boards rather than aggregators, and enforces location constraints in code instead of asking a model to remember them.

taper — a running simulator grounded in exercise-physiology research. Oregon Trail for runners, built on a decade of my own Strava data.

waxcast — kick and glide wax calls for Anchorage ski trails, with the reasoning attached. Every wax chart online asks for the air temperature, which is the wrong input twice over: wax responds to snow temperature, and snow that has thawed and refrozen for three days behaves nothing like fresh snow at the same reading. This works from the three-day history instead. No build step, no server, no API key.

mtg_ai — sealed-deck builder that reads your Arena pool straight from the game log, scores all ten color pairs on real 17Lands win-rate data, and checks the manabase against Frank Karsten's castability thresholds. Python API, React front end.

covered-call-copilot — turns a brokerage export into one ranked options-income plan. Runs on localhost; nothing leaves the machine.

Reach me at nathaniel.knapp@gmail.com or on LinkedIn.