CS @ UT Dallas '28 · building AI systems that have to work in the real world

I'm a Computer Science student at UT Dallas (minor in Math) who spends most of his time on the messy parts of AI systems — the state machines, the crash-safety, the latency budgets, the thing that breaks when a real camera feed hits a model trained on grayscale data.

• 🔬 Undergraduate Researcher @ UTD URAP (advised by Prof. Wei Yang) — studying the root causes of non-determinism in LLM inference. Curated 200+ documented cases and built harnesses that replay divergent runs.
• 🏆 2nd Place out of 350+ at HackUTD 2025 (PNC Track) · Dean's List
• 💬 Ask me about edge ML, agent orchestration, RAG that actually cites its sources, or why your WebSocket server is leaking hole cards.

🗣️ Holler — a multichannel AI agent harness

Unifies WhatsApp, Telegram, and Twilio into one persistent, event-driven agent runtime with Gmail, Outlook, Drive, and Calendar exposed as callable tools. Hit 91% task completion on a 100-workflow eval suite while cutting LLM calls 28% and cost 24% vs. a multi-agent baseline. Destructive tool calls are crash-safe: each proposal commits to a single-use pending-action record in WAL-mode SQLite and is consumed atomically, so a replayed turn drops the action instead of double-sending. Python FastAPI TypeScript SQLite Twilio Composio RAG

♠️ Live Multiplayer Poker — authoritative real-time game server

360 concurrent players across 60 tables at 1,871 actions/sec with 0.3 ms p95 action-to-broadcast latency. Deterministic seeded-RNG engine with chip-conservation invariants, proven across 3,000 fuzzed hands and 106 tests. Per-viewer state projection plus a recursive payload scanner in CI asserts zero hole-card leakage across every player, spectator, and history view. React TypeScript Node.js WebSockets Supabase Vercel

📋 AutoPM — AI product-management copilot · 2nd Place, HackUTD 2025

Orchestrates 8 LangGraph agents across a 7-step workflow — research, user stories, RICE prioritization, wireframing, Jira tickets — with Zod-validated structured outputs across 3 third-party APIs. Cuts an estimated 10+ hours of manual PM work per cycle. Next.js TypeScript LangGraph Gemini

📚 Research Mate — RAG over papers that cites its page

Two-node LangGraph pipeline retrieving the top 10 pages from a hybrid ColQwen2 index over Qdrant and FAISS, answering on Groq Llama 3.3 70B constrained to cite its source page. Ingestion is crash-safe and idempotent via a Postgres state machine and deterministic sha256(pdf_id | page) vector IDs. Python LangGraph ColQwen2 Qdrant PostgreSQL Gemini

🏀 NetGainNBA — playoff prediction + GM toolkit

88% playoff-prediction accuracy on 2020–2024 data from a 7-model ensemble across 89 features, with live win-probability tracking and K-Means player archetypes. React Flask XGBoost scikit-learn

Languages

AI / ML

Web & Backend

Cloud, Edge & DevOps

Officer at UT Dallas's Artificial Intelligence Society, where I help run HackAI — the university's annual AI hackathon — and lead technical workshops. Also Product Manager for TuneTrend in the UTD AI Mentorship Program, directing a 5-person team building a genre-aware music popularity predictor.

Open to Summer 2027 SWE internships — say hi.