Senior AI Engineer — RAG, multi-agent orchestration, LLM evaluation, and MLOps. 8+ years shipping production LLM and ML systems in regulated domains (pharma / FDA, banking / fintech, B2B), plus the unglamorous data engineering that keeps them running in production instead of just in a demo.

I care about one thing in AI engineering: claims you can check. The systems below run end to end; the benchmarks regenerate their own numbers. No API keys, mystery services, or copied paper results. The same discipline spans LLM infrastructure and agents through bioinformatics, genomics, vision, and audio.

🟢 Open to senior / staff / lead AI & ML engineering roles (IC or founding) and AI-product roles — remote, or relocation for the right team. → adoghri.com

Real systems shipped in regulated and B2B environments. Proprietary, not open source, but the discipline behind them is exactly what the repos below are built to demonstrate.

These are not notebook demos. Each one has a usable interface, a backend or CLI, persistence where the product needs it, tests, and a documented local or containerized run path.

Ten local-first systems I wanted to exist: tools that decide what to trust across biology, agents, media, and audio. Each ships a CLI, JSON API, browser workbench, Docker image, tests, and an explicit boundary around what its result does not prove.

Ten useful experiments at the edge of current ML engineering. Every result below comes from the committed fixture or an actual generated media artifact, not from the source paper.

Six focused repos that isolate the hard parts of shipping LLM systems: cost, safety, hallucinations, memory, evals, and injection defense.

Current research reduced to one mechanism, one baseline, and one number you can rerun. Every repo is MIT licensed, standard-library Python, and tested on 3.9, 3.11, and 3.13.

The same standard (reproducible benchmark, tests, zero dependencies) applied across the stack. Grouped so you can jump to what you care about.

LLM inference and cost. speculabench (speculative decoding math, 1.4x to 2.9x — the optimal draft length shifts once real bursty agreement patterns replace an independent-draw assumption), kvsqueeze (KV-cache eviction, H2O and StreamingLLM style — the original policy's "win" was freezing the cache entirely, scoring 0% on recent-token recall; fixed), contextpack (prompt compression with a recall check — the safe-compression knee point only held when the source text used digit numerals; fixed for prose-spelled numbers too).

Output reliability. structstream (JSON repair, 7% → 100% recovery on the bundled corpus, 0% on a common failure mode it originally never tested — now fixed), rubricagent (LLM-as-judge rubrics learned from outcomes, AUC 0.77 → 1.00, confirmed on a disjoint holdout), taggate (confidence-gated tagging in TypeScript).

RAG done honestly. chunklab (chunking strategies, measured — the 30-point margin on the bundled document nearly disappears on a second, independent one), clarifyrag (asks clarifying questions only when evidence disagrees — auditing its own circular benchmark surfaced a real stopword-filtering bug, fixed, natural-question accuracy 3/7 → 7/7), tablextract (tables out of PDFs with cited cells — found and fixed silently-dropped rows and fabricated citations, recall 71% → 100%).

Agents that stop. toolrouter (tool selection that abstains on ties, though it also abstains on 10/15 real natural-language queries — fixed), agentbudget (loop detection catching stalls that step limits miss, plus a blind spot of its own it didn't originally catch), debatekit (multiagent debate: the panel effect, 57% → 79%, is real; the extra lift from debate rounds specifically never clears one standard error at proper sample size).

ML beyond LLMs. churnfm (drift-triggered retraining), orthoshift (double ML for causal effects — a multi-seed check found plain adjustment quietly beats it as often as it loses), fedcal (non-IID federated learning — a single-seed calibration win reverses to a net loss across 60 seeds), riskbandit (conformal risk-controlled bandits — the quantile formula undercovered its own stated 90% target by 1-1.5pt, fixed), chronopatch (conformal forecasting — the gain holds on the paper's exact seasonality, drops to single digits on a different but equally realistic series shape), graphpulse (graph anomaly scoring — most of the published AUC traced back to reading the ground-truth label directly; the honest label-free score is lower but still clearly ahead of degree alone), tabflowmini (synthetic tabular data with an audit — the "fit" churn model was the generator's own coefficients copy-pasted, now actually fit), proteinmask (masked protein-like infilling, honestly toy — its "random baseline" was a rigged formula tuned to never match, fixed to a real uniform draw), pendulumlab (CEM control from scratch — beats a fair grid-search baseline, not just an under-tuned guess), motifdiff (symbolic music that grades itself — isolated the weighting's real contribution once two confounds were removed), connectpuct (PUCT Connect Four you can play — a perfect record vs. weak baselines drops to ~55% against real depth-3 minimax, a fair fight).