Quality engineer (ASQ-CQE, 15+ years in regulated hardware manufacturing) who deploys and operates private LLM infrastructure: multi-node serving, evaluation gates, silent-failure monitoring and disaster recovery, run with the same discipline as a production line.
I build with AI coding agents (Claude Code, Codex, GitHub Copilot CLI). I set the architecture, constraints and acceptance criteria, then review, test, deploy and operate what they write.
| Repo | What it shows |
|---|---|
| LLM-chatbot | Question answering over technical manuals in which every claim carries a verbatim quote, checked and highlighted on its source page: PaddleOCR-VL parsing with page coordinates, tables as data, hybrid retrieval with a Qwen3 reranker, any OpenAI-compatible model. A pre-registered held-out test with blind human grading: 80% of answers correct and supported against 42% for v1 (the pipeline and a larger model together). Try it in the browser: https://alex8642.github.io/LLM-chatbot/ |
| onprem-llm-stack | Two NVIDIA GB10 nodes serving large quantized models (MiniMax M2.7, Qwen3.5-122B, Qwen3.8-Flash-Next) tensor-parallel on vLLM / SGLang / Atlas, behind a LiteLLM gateway and OpenWebUI with SSO. Includes an A/B evaluation harness that caught a faster recipe failing long-context retrieval, a watchdog for a silent NaN-logits fault (detect → contain → correct → escalate), an Ansible rebuild with a written DR runbook, a measured trial of Cohere Command A+ (218B MoE, TP=2), and a tool-call eval of the Home Assistant voice agent |
| audio2report | Dual-mic meeting recordings → deduplicated transcript → LLM audit report, fully local (WhisperX, pluggable local LLM backends, CPU/GPU Docker images, CI) |
| lights-out | A self-hosted model that troubleshoots the network when the internet (and every cloud assistant) is down, tested on replayed incidents. It never gets a shell: read-only commands pass a policy enforced in code, and any fix is a closed list of scripts whose risk class decides whether it acts, asks me, or is refused. Scored on diagnosis, restraint and evidence, with success rates and 95% intervals per batch and energy per correct diagnosis measured at the wall plug. First batch: 0/5 correct at temperature 0 vs 4/5 at 0.7 on a two-cause DNS fault; no unsafe command proposed so far. Run as a reviewed-PR project: issues, ADRs, CI, changelog fragments |
| judge-msa | A gauge study (Gage R&R) for LLM judges: repeatability, agreement between judges, bias against a human reference, and the smallest difference a judge can resolve, validated on simulated data with planted truth. Worked study on LLM-chatbot's judges with a protocol committed before the runs: identical verdicts at temperature 0, yet about 7% change when the key facts are reordered; each judge resolves differences only above 19–21 points; an unrecorded context-window default changed verdicts on long inputs |
| Repo | What it shows | Headline |
|---|---|---|
| wafer-defect-classifier | 9-class WM-811K wafer maps: focal loss + CBAM, temperature-scaled calibration, Grad-CAM++ | macro-F1 0.92, ECE 0.003 |
| wafer-ssl | SimCLR pretraining on 638k unlabeled maps + 4-seed ensemble, with an ablation attributing the gain | macro-F1 0.94 |
| wafer-mixed | Multi-label mixed-defect classification (MixedWM38), transfer study, cost-weighted per-label thresholds | macro-F1 0.98 |
| wafer-rootcause | Simulated MES in DuckDB; SQL commonality analysis (z-tests + BH-FDR) recovers planted equipment faults | precision@1 ~0.9–1.0 |
| wafer-deploy | FastAPI serving with a bit-for-bit parity test, three drift monitors, hysteresis retrain trigger, Prometheus + Grafana | scored detection latency and false-alarm rate |
Negative results stay in the docs. For example, a pseudo-labeling regression was rolled back and a loss-function bug is written up as a post-mortem.



