Selected work
Three production systems. Measured, deployed, observable.
Each project ships with eval rigor — golden datasets, regression suites, and cost dashboards. Live links below; deeper walkthroughs on request.
Sentinel Vision
Custom 96-dim color histogram embeddings with cosine distance — AUROC 1.0 at 1.4 ms/frame, outperforming CLIP (0.891) and ResNet (0.923). Agent-harness eval suite benchmarked embedding variants systematically and cut inference overhead 28%. Live deployment with Docker Compose, Nginx, Let's Encrypt and a real-time confidence dashboard.
- 1.0
- AUROC
- 1.4ms
- per frame
- -28%
- inference overhead
DriftWatch
Multi-provider LLM evaluation (GPT-4o, GPT-3.5, Claude) with Agent Harness Engineering: RAGAS-style scoring at faithfulness 0.91, answer relevancy 0.88 on a 100-query eval set. Token usage reduced ~35% ($220 → $143/month) via harness-enforced scope constraints and semantic caching. Statistical drift detection (EWMA, PSI, KS) with a model-arena leaderboard.
- 0.91
- faithfulness
- -35%
- token spend
- <500MB
- docker image
Enterprise Multi-Agent Research Assistant
Planner agent decomposes queries → specialized sub-agents (RAG retrieval, web search, code execution) → Synthesizer produces cited answers. LLM-as-Judge eval harness scores faithfulness 0.87 and answer relevancy 0.84 on a 50-query golden dataset. Cost-aware model routing and harness-enforced token budgets cut per-query cost ~62%. Deployed on DigitalOcean with Prometheus/Grafana observability and CI/CD regression on every push.
- -62%
- per-query cost
- 0.87
- faithfulness
- 0.84
- answer relevancy