§ Précis
I work the hard parts of LLM products — retrieval, evals, observability, cost routing — and ship them to production with measurable wins. Currently at AriesView; M.S. Machine Learning at Stevens, class of '26.
Resume ↓
§ 01 · About
Most teams know how to call an API. The hard parts are retrieval that doesn't lie, evals that don't drift, and bills that don't grow exponentially.
That's where I work. Production RAG with hybrid + graph retrieval, agentic systems wired up via MCP, semantic routers that cut cloud spend by orders of magnitude, and CI/CD gates that block hallucinations before they ship.
Open-source Model Context Protocol server exposing GitHub and internal CRUD APIs as tools to LLM agents. GitHub OAuth2 auth, token-bucket rate limiting via Redis (500 req/min). One-command Docker Compose deployment. 558 req/s throughput, 3.7ms health check latency, p50 API latency ~5ms.
Heuristic prompt complexity classifier (0.06ms avg, 0.17ms p99 — no model inference) routing 79% of queries to local Llama 3.2 (vLLM) and complex tasks to GPT-4o. 93.4% routing accuracy across 81-query benchmark. Streamlit FinOps dashboard tracks per-query cost and routing decisions. 68.1% reduction in cloud inference spend ($0.4835 → $0.1542).
CI/CD eval gate that runs a RAG chain through three DeepEval metrics (Faithfulness, Answer Relevancy, Contextual Precision) on every pull request using Claude Haiku as judge — blocks merges when any metric falls below 95%. LangSmith traces every retriever call and generation with inputs, outputs, latency, and token counts. Built to catch silent hallucination regressions from prompt or model changes before they ship.
Autonomous multi-agent equity analyst ingesting SEC 10-K filings, real-time prices, and earnings transcripts. LangGraph StateGraph orchestrates five sub-agents: filing parser, DCF valuation, FinBERT sentiment, moat assessment, and memo synthesis. Neo4j knowledge graph for cross-company comparative queries. Full pipeline traced end-to-end in LangSmith. Generates cited Buy/Hold/Sell memos with confidence scores in under 30 seconds.
Hackathon project. MCP server that reads a patient's FHIR R4 bundle, queries ClinicalTrials.gov for recruiting trials, and uses Claude to reason plain-English inclusion/exclusion criteria against the actual clinical profile — returning a ranked, scored, explained list directly inside the clinician's workflow. Targets the discovery gap behind the 3–5% trial enrollment rate despite ~50% medical eligibility. SHARP context propagation auto-injects patient context from the EHR. Patient-facing summaries in 7 languages.
End-to-end ML pipeline predicting 30-day diabetic patient readmissions across 101,766 encounters (UCI). Four-phase methodology: preprocessing + PCA, K-Means and hierarchical clustering to surface patient risk groups, SMOTE-balanced classification, and full evaluation. Six classifiers benchmarked — Logistic Regression recommended: F1 0.2488, ROC-AUC 0.629, Recall 52.8%. Random Forest hit 88.8% accuracy but caught only 3.2% of true readmissions, illustrating why recall matters most in clinical settings. Stevens MIS 637.
§ Education
§ Certifications