I build LLM systems, then spend just as long proving they work.

Built a temporal evaluation environment — a small discrete limit-order market with exact rational transition probabilities — so forecasts are scored against the exactly computed conditional law instead of one sampled outcome, removing sampling noise from the metric itself.

Computed the target by Bayesian filtering and exhaustive enumeration in rational arithmetic, checked against two independent solvers and a Monte-Carlo simulator, then generated 300,000+ paired natural-language and structured problem instances. Reference learners trained to convergence on a five-rung difficulty ladder pin down the accuracy ceiling a model can actually reach — one evaluated model closes the gap from many times the baseline down to that ceiling at roughly 3,000–9,000 reasoning tokens.

Designed an embargo-period protocol — all evaluation data dated after the teacher model's training cut-off — with multi-seed replication and matched-stack, label-shuffle, and threshold-sweep controls, to separate genuine model behavior from frontier-API response variance. Found that prompt-level differences for an LLM crypto-trading agent are mostly seed noise: none of four paired treatment effects had a 95% confidence interval excluding zero.

Fine-tuned Qwen3-8B with supervision-only LoRA on pre-embargo frontier-model decisions. One risk behavior — avoiding long exposure on crash days — survived replication and transferred to the small open model.

Co-built ACE, an internal Automated Compliance Engine on the Claude Agent SDK that drives GitHub, AWS, and Playwright browser automation to collect ISMAP and ISO/IEC 27001 audit evidence. Implemented 40 ISMAP and 12 ISO controls end to end, a dashboard over a REST + WebSocket API, per-run cost monitoring, and multi-model routing to keep spend down.

Cut evidence collection for a control set from about five working days to under half a day. Internal teams tested it against real ISO evidence and reported edge cases that were fixed in turn.

Deployed a fully on-premise RAG knowledge base on an NVIDIA A800 80GB server — DeepSeek-R1-Distill-Qwen-14B served with vLLM, plus Qwen3-Embedding, Docling ingestion, and Open-WebUI, all containerized with no data leaving the local network. Ran functional, performance, and concurrent-load stress tests, then wrote the evaluation report: recommended vLLM over Ollama on throughput, security, and OpenAI-API compatibility.

Lead the student team that keeps the department's shared Linux workstations running — server operations, user support, environment maintenance, and Python / shell automation for routine tasks.

Real transaction data, trend analysis, and a Chrome extension for in-browser reporting.

Carry and near-settlement "snipe" modes, with dual-margin monitoring sized to survive ±30% moves on either leg independently.

Parses unstructured Telegram channel messages with a local Ollama model into structured Binance Futures orders — fully on-device inference, zero cloud. The project that pulled me into quantitative trading.

Diagnoses why AI assistants recommend a competitor's product, splitting the cause into information gap vs. product gap with a before/after validation loop.