My Work

Things I've built — agentic AI platforms, data infrastructure, and the engineering work behind them.

Projects

Agentic AI portfolio and stock analysis platform

Full-stack investment intelligence platform (React/TypeScript + FastAPI) deployed on Vercel and Render, with Neon Postgres persistence, live market data, and dual-LLM routing with automatic failover.
A 7-agent analysis pipeline — Ticker Intake, Market Data, Forecast, Backtest, Portfolio, Memo Writer, Compliance — where deterministic code computes every number and an LLM-narrated, compliance-validated layer explains it, guarded by regex rules against financial-advice language.
Real external integrations: Plaid brokerage sync (raw REST, no SDK) for live holdings import, and a Google A2A-compliant agent card that exposes the platform's skills for agent-to-agent discovery.
Production infrastructure — rate limiting, structured request-ID logging, Sentry-ready error tracking, a 30-test CI suite (pytest + Vitest) gating every deploy, and a replayable analysis history backed by persisted agent traces.

Adaptive LLM orchestration and prompt optimization platform

Provider-agnostic AI workspace built with Node.js, vanilla JavaScript, and Vercel, supporting automatic LLM failover, SSE-streamed progress, request cancellation, and provider-reported token and cost tracking.
An eight-module orchestration engine moving each request through secret redaction, adaptive route selection, prompt construction, typed handoff generation, provider execution, optional verification, and usage accounting behind one unified API.
A hub-and-spoke agent workflow where intake, contract, execution, and verification stages exchange compact validated state instead of full conversation transcripts, while emitting trace IDs, agent actions, durations, success states, and failure details.
Production safeguards including Zod validation, credential-pattern redaction, SSRF protection, IP-based rate limiting, CSP/HSTS headers, and path-safe static serving, tested automatically across Node.js 20 and 22.
A reusable Manifest V3 Chrome extension with Gemini and ChatGPT adapters, plus Graphify — an interactive code map visualizing files, functions, dependencies, and agent information flow across the open-source system.

Jun 2026 — Present

Software Engineering Intern — Platform Generative AI · Bethesda, MD

Built an A2A-enabled multi-agent AI platform from Marriott's Java starter kit, evolving a single-service MCP/RAG prototype into a coordinator-led runtime with specialist agents, routing rules, gRPC communication, streaming APIs, memory, monitoring, and dashboard controls.
Expanded the platform across 5 backend modules and a React/Vite dashboard, growing the system from 36 to 250 Java classes and from 6 to 105 mapped endpoints across coordinator, analytical, news, mock-agent, and knowledge-layer services.
Integrated PostgreSQL/pgvector RAG over 147 hotel inventory files using 1024-dimensional embeddings, scoped property retrieval, top-k search, and chunked JSON ingestion to ground agent responses in Marriott property data.
Added scalability and reliability features including A2A agent discovery, gRPC with HTTP fallback, Redis-backed session memory, configurable routing, multi-level caching, health probes, audit logs, and performance metrics.

Jun 2024 — Aug 2024

Research Intern — Data Analytics (Immuno-Oncology) · Washington, DC

Analyzed high-dimensional flow cytometry data to identify immune cell response patterns across complex datasets.
Cleaned and validated experimental data with Excel and Python to support reproducible statistical analysis.
Presented data-driven findings at weekly research symposiums and worked with researchers to interpret results and recommend next-step analyses.

Contributions live on GitHub.