Andrew Woods

Projects

Built. Tested. Plotted.

Explore the systems through their results.

Intent classification / accuracy
Local specialist85.58%
Routed hybrid88.08%
LLM on every request92.18%
050100%

Same 3,080 test cases. Three ways to answer.

ML systems · 2026

Let the specialist handle the easy ones.

88.08% accuracy. Only 8.70% of requests sent to an LLM. Explore where the hybrid sits between cost and quality.

Explore intent routing →
77 intents268 / 3,080 routed to LLMMiniLM + logistic regression
Retrieval / nDCG@10
0.80.40 Hybrid RRF: 0.7216 nDCG@10 Reranker: 0.6879 nDCG@10 0.7216 0.6879 Hybrid+ Reranker
Hybrid RRF+ Reranker

Retrieval · Evaluation

The reranker made it worse.

Lower retrieval quality on two of three datasets. Roughly 70 times the latency on SciFact. The charts tell the story.

Explore the benchmark charts →
−0.0337 nDCG@10 on SciFact34.5K documents1,623 judged queries
OmicsAI / deployment architecture
Caddy · HTTPS ingress
PortalReact / Nginx
ResearchGradio
ResearchStreamlit

Four Docker services on AWS EC2. One entry point.

Software engineering · OmicsAI Labs

From research code to a working platform.

Three applications behind one ingress, with isolated ports, persistent state, and a six-page research portal.

Visit the research demo ↗