Senior AI Engineer. 8+ years building ML across bioinformatics, genomics, vision, video, sound, LLM systems, and 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. Every open source repo below ships a benchmark that reproduces locally, with no API keys and no asterisk. If a number is in a README, a command in that README regenerates it.
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.
| Project | Description |
|---|---|
| LLM-Powered Regulatory Submission | Automated FDA submission process using 7+ LLM pipelines, reducing errors by 60%. |
| AI-Driven RAG System | NLP-powered retrieval-augmented generation system for pharmaceutical documents. |
| Streaming Data Platform | Real-time event-driven ML pipeline on Kafka, Spark, and Flink. |
| Biomedical Text Enrichment | Transformer models enriching research papers with biomedical insights. |
Six repos that cover the full surface of shipping LLM systems: cost, safety, hallucinations, memory, evals, and injection defense.
| Repo | What it proves | Result |
|---|---|---|
| vllm-cost-router | Complexity-based routing, caching, and batching in front of vLLM | 73% lower cost and p95 latency vs. always-large-model serving |
| guardrail-gate | PII redaction + citation grounding + rate limiting in one pass | 100% precision/recall on structured PII, 83% grounded-vs-hallucinated accuracy |
| semanticentropy | Hallucination detection via semantic entropy (Farquhar et al., Nature 2024) | Consistent answers score 0.08, hallucinations 0.90. No labels, no judge model |
| injectguard | Prompt injection and jailbreak detection with explainable verdicts | 100% precision and recall on a red-team corpus of attacks and lookalikes |
| agentmem | Bounded, self-consolidating long-term memory for LLM agents | Salience-gated writes, decay-aware retrieval, a hard budget it actually respects |
| citebench | What reranking actually buys you in citation-grounded RAG | Citation precision 62% → 88% on an adversarial benchmark |
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.
| Theme | Repo | What it proves | Reproduced result |
|---|---|---|---|
| Bioinformatics | cellcontext | Cellular context matters for perturbation response | MAE 0.2699 → 0.1813 |
| Bioinformatics | foldcontact | Contact constraints keep protein infilling fold-aware | Satisfaction 34.5% → 100% |
| Genomics | pangraphmap | Pangenome paths recover structural-variant reads a linear reference loses | 6/25 → 25/25 mapped |
| Genomics | methyloadapt | Conserved motifs transfer when target-species labels run out | Accuracy 50% → 100% |
| Machine learning | driftfilter | Forward-only prototypes follow gradual deployment drift | Accuracy 77.9% → 100% |
| Machine learning | taskrouter | Routing preserves specialists that static merging averages away | MAE 1.379 → 0.000 |
| Vision + video | distractrack | Motion and identity memory prevent swaps at object crossings | Accuracy 73.8% → 100% |
| Vision + video | d3video | Synthetic motion artifacts show up in the difference of differences | Accuracy 50% → 100% |
| Sound + music | restem | A separator can improve itself over multiple inference steps | SNR 8.78 → 36.52 dB |
| Sound + music | binauralbench | A clean stem still fails if it collapses the spatial cue | ILD error reduced 98.3% |
The same standard (reproducible benchmark, tests, CI, 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), kvsqueeze (KV-cache eviction, H2O and StreamingLLM style), contextpack (prompt compression with a recall check).
Output reliability. structstream (JSON repair, 7% → 100% recovery), rubricagent (LLM-as-judge rubrics learned from outcomes, AUC 0.77 → 1.00), taggate (confidence-gated tagging in TypeScript).
RAG done honestly. chunklab (chunking strategies, measured), clarifyrag (asks clarifying questions only when evidence disagrees), tablextract (tables out of PDFs with cited cells).
Agents that stop. toolrouter (tool selection that abstains on ties), agentbudget (loop detection catching stalls that step limits miss), debatekit (multiagent debate, 57% → 80%).
ML beyond LLMs. churnfm (drift-triggered retraining), orthoshift (double ML for causal effects), fedcal (non-IID federated learning), riskbandit (conformal risk-controlled bandits), chronopatch (conformal forecasting), graphpulse (graph anomaly scoring), tabflowmini (synthetic tabular data with an audit), proteinmask (masked protein-like infilling, honestly toy), pendulumlab (CEM control from scratch), motifdiff (symbolic music that grades itself), connectpuct (PUCT Connect Four you can play).
| Tool | What it does |
|---|---|
| VectorMorph | SVG to animated WebP/GIF with frame blending. Built because Figma's export flaked one time too many. |
| ATSProofResume | Tailors your resume to a job posting without inventing experience. The free version of a $29/month industry. |
| bookconverter | Local EPUB/MOBI/PDF conversion. No upload limits, no queues, no well-hidden subscription. |
