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.
| 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. |
These are not notebook demos. Each one has a usable interface, a backend or CLI, persistence where the product needs it, tests, CI, and a documented local or containerized run path.
| Repo | What ships | Check it |
|---|---|---|
| spanjudge | OTLP agent-trace receiver, SQLite store, operational dashboard, JSON API, CLI, Docker image, and regression policy gate | Six spans rebuild three traces; the fixture passes latency, error, cost, and eval release limits |
| vrsbridge | VCF parser, GA4GH VRS 1.3 translator, equivalence engine, browser workbench, JSON API, CLI, and Docker image | Four VCF encodings collapse into two normalized molecular variants |
| leterminale | Browser terminal, encrypted client-side workspace, import/export, keyboard workflow, and static deployment | Open the live product and use it without an account or server-side document store |
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, CI, and an explicit boundary around what its result does not prove.
| Theme | Repo | What it checks | Demo result |
|---|---|---|---|
| Spatial biology | spatialniche | Cell-type neighborhood enrichment against a permuted null | Tumor neighborhood z = 5.10 across 250 permutations |
| Protein structure | structuregrade | Per-residue pLDDT bands and inter-residue geometry | Grade C, mean pLDDT 71.88, two residues below 50 |
| CRISPR genomics | crisprradar | Both-strand SpCas9 NGG mismatch and seed risk | 1 exact + 2 off-target sites across 149 bases |
| Clinical genomics | phenopacketlint | GA4GH Phenopacket semantic exchange readiness | Three phenotype assertions, quality score 100 |
| Agent security | mcpinterlock | MCP tool authority, paths, approvals, secrets, and SSRF | One call denied with 2 independent violations |
| ML privacy | unlearnaudit | Unlearning leakage against retained utility | Membership AUC 1.000 → 0.481, accuracy stays 0.988 |
| Video privacy | videoprivacy | Tracked redaction through detector gaps | 10 regions, two identities, one gap filled |
| Media provenance | manifestlens | C2PA ingredients, actions, signatures, and hard binding | One ingredient, three actions, valid hard binding |
| Audio delivery | loudnessgate | EBU R128 loudness, range, peak, and normalization | -18.4 LUFS, with a measured +2.4 dB correction |
| Music libraries | audiocatalog | Chromaprint duplicates across renamed or transcoded files | One duplicate at 98.65% similarity |
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.
| Theme | Repo | End-to-end job | Demo result |
|---|---|---|---|
| Genomics | strandshift | Audit sequence models across strand and window transforms | 18 views, prediction range 0.7309 |
| Rare disease | phenorank | Rank HPO disease candidates and test stability | Top score 0.6334, stable 3/3 |
| ML security | tensorwarden | Scan real checkpoint containers without loading them | Safe artifact accepted; 2 quarantined |
| LLM privacy | cacheisolate | Reproduce and stop cross-tenant cache timing leakage | 86 ms oracle removed with useful reuse retained |
| RAG security | ragpoisonbench | Inject corpus poison and test provenance quarantine | Recall 1.0 → 0.0 → 1.0 |
| 3D vision | splatgrade | Parse and grade Gaussian Splat PLY artifacts | Grade D / 49, every defect localized |
| Generated video | physicsvideo | Render MP4 scenarios and audit physical consistency | 4/4 faults, 0 false positives |
| Multimodal media | avsyncdoctor | Detect and repair offset plus clock drift | 120 ms, 5,000 ppm, repaired below 40 ms |
| Audio engineering | codecguard | Run real codec round trips behind a CI quality gate | Clean build passes; degraded build fails 4/4 gates |
| Robot safety | safepathshield | Project policy actions onto a barrier constraint | Collision steps 39 → 0, goal still reached |
Six focused repos that isolate the hard parts 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 |
|---|---|
| leterminale | A private, keyboard-first writing terminal with encrypted local workspaces and no account requirement. |
| 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. |