AI Engineer · LLM Systems · Applied ML
I work across the part of AI that starts after the demo: model adaptation, retrieval, agents, evaluation, serving, observability, and failure analysis.
| System | Engineering problem | Evidence |
|---|---|---|
| OpsVerse · demo | Post-train Qwen3-4B, retrieve with BGE + BM25 + RRF, rerank, ground citations, gate releases with evals, serve with vLLM | MRR@10 0.705 · 1,914 graded judgments · 245 regression tests · ~6× lower cost/response |
| AxonFDE · demo | Fuse telemetry, orders, and shipping docs; predict breaches; reconcile conflicts; require human approval; audit every action | 49 min median lead time · 23/23 planted conflicts · 0 unauthorized actions / 50 tests · 4.92 s decision cycle |
| Pro2Pro · demo | Checkpointed LangGraph + MCP orchestration with resumable state, human approval, guardrails, tracing, and model fallback | 78 evaluation tests |
| FIFA2026MLOps · demo | Elo + gradient boosting + Monte Carlo forecasting with automated training, validation, deployment, lineage, and drift monitoring | Airflow · MLflow · Evidently · OpenTelemetry |
I look for numbers that are quietly wrong in LLM serving and energy tooling, reproduce them, and fix them in the open.
| Project | What was wrong | Where it stands |
|---|---|---|
| AIBrix routing | Least-latency scored pods by a lifetime average, so a pod that had slowed down kept winning traffic for about 10 minutes after 100,000 requests. A one-minute window cuts that to 15 seconds. | Issue #3297 · PR #3327: open, reviewed by the maintainer, follow-up pushed for his notes |
| AIBrix metrics | An empty histogram was exported as a real percentile. | PR #3290: approved, checks green |
| AIBrix batch | One NaN latency in a response leaves the adaptive concurrency controller at full load while the engine slows 50x. |
Issue #3339: accepted by maintainers |
| vLLM | Reproduced a streaming bug and validated another author's fix on a GPU: 18 of 640 requests broken before, 0 of 640 after. | Validation on #62726 |
| CodeCarbon | Country carbon intensities were 2 to 4 years old. A script now rebuilds them from the maintained dataset. | PR #1468: open |
| GuideLLM · Ray | Missing token timings counted as zero latency; a crash and lost prompts in Ray's vLLM batch stage. | #1406 · #67460: open |
llm-bench-wire-audit: a logging proxy that checks what vllm bench serve, GuideLLM and AIPerf report against what actually crossed the wire. Where definitions match they agree with the wire within a few percent. Under an injected 300 ms stall, one tool's p99 inter-token latency reads 311 ms and another's reads 16.6 ms, because one averages per request and one does not.
|
Research with Prof. Kenneth W. Regan Compare a Regan-inspired statistical reconstruction, Qwen3-8B, and Maia-3 on the same chess decisions using Stockfish MultiPV and controlled engine-assisted substitutions. Measured result
|
Research with Prof. Haipeng Cai Systematizing failure modes of LLM/VLM-driven physical agents across prompt/tool attacks, adversarial perception, leakage, and unsafe actions. Current direction LLM-assisted vulnerability localization → patch proposal → sandbox regression test. |
MODELS PyTorch · Transformers · Qwen · TRL · PEFT · QLoRA · SFT · DPO
RETRIEVAL RAG · BGE · BM25 · RRF · reranking · Qdrant
AGENTS LangGraph · LangChain · MCP · tool use · human-in-the-loop
EVALUATION benchmarks · regression tests · paired bootstrap · guardrails · red teaming
SERVING FastAPI · vLLM · REST APIs · Docker · Kubernetes · AWS
OBSERVABILITY Langfuse · OpenTelemetry · MLflow · Evidently · CI/CD
- GWO-Based Fed-UNet-CNN Model for Leukocytes Classification Across Developmental Stages
- The Future of Smart Factories: Real-Time Automation and Big Data Analytics
