Intelligence is only the beginning. Making it useful is the work.
I build across the stack, from retrieval and model inference to APIs and the interfaces people use. The interesting part is making all of it work together.
A document-intelligence system that turns 400–800-page loan agreements into 22 source-linked fields. Retrieval, reranking, validation and correction rules make the output traceable, not just plausible.
Built and handed off for secure, non-production Azure deployment. Review time fell from 30–60 minutes to under 5; executives estimated 300+ hours of potential savings.
Document intelligenceSource-linked output
From agreement to evidence.
Input400–800 pages
Structured output22 fields
RetrievalFAISS + reranking
ValidationAnswer-key agreement
Correction layerReusable JSON rules
System overview, not a screenshot of confidential employer data.
Co-developed · Cloudflare Workers / Hono
The right listing. Not yesterday’s.
A deployed real-estate search and alerting platform with public read APIs, protected writes and an auditable trail of new listings, price changes and status changes.
A change worth knowing.Auditable price-change events
Freshness72-hour window
Old data has an expiry.Stale listings are explicitly identified
Illustrative workflow, not live property data.
Realtor.com Prediction Capstone
A team capstone using Python, logistic regression and XGBoost to study future user inquiries and inactive-member reactivation as separate prediction problems. Time-ordered features, class-imbalance-aware evaluation and reproducible modeling workflows. Sponsor data remains private under NDA.
Embedding Inference Service
A self-hosted FastAPI batch embedding API with CPU/GPU execution, model caching, health checks, telemetry and load tests. Benchmarks document cost–latency tradeoffs against hosted inference.
A reproducible PyTorch pipeline over 100K+ ECG recordings: 0.821 test AUROC, 0.791 AUPRC and 0.018 calibration error. Research only; not clinically validated.
A React chatbot over a proprietary insurance API, making operational data accessible to engineering and transactions teams through natural language. Completed a non-production handoff.
Flask inference for a 12-layer TensorFlow Conv1D classifier across 20+ investment intents and 2,000+ labeled examples, with 89% held-out accuracy and an optional LLM response path.
From enterprise document workflows to inference services and full-stack applications.
Revantage (Blackstone)
AI Engineering Intern
Built and handed off document intelligence and an API-backed insurance assistant for secure, non-production Azure deployment. Python, FastAPI, React, FAISS and PostgreSQL.
Glynac AI
Initial Member of Technical Staff
Built investment-intent inference and LangChain/Qdrant semantic retrieval over 100+ investment documents, returning source-grounded evidence through vector search.
Springer Capital
Data Science Intern
Engineered a workforce-intelligence application with Next.js, TypeScript, Prisma and Azure PostgreSQL to centralize analysis and surface sentiment-derived employee-risk signals.
JPMorgan Chase
Software Engineer Intern
Built a Java/React banking dashboard backed by Cassandra, with editable account workflows and automated interest calculations for internal users.
AI evaluation
Good answers stand up to hard questions.
Mercor · Reviewer
Promoted from Senior AI Writer to Reviewer. Review AI training tasks and workflow demonstrations for accuracy, completeness and adherence to task requirements under NDA.
Mercor · Senior AI Writer
Authored multi-step tasks and recorded end-to-end workflow demonstrations for AI training, translating complex objectives into explicit action sequences and reproducible examples under NDA.
Handshake AI · LLM Evaluator
Authored challenging, document-grounded prompts to expose LLM reasoning and retrieval failures; developed reference solutions and refined grading rubrics to assess response accuracy and completeness under NDA.
Mercor and Handshake AI experience consists of contract AI training and evaluation work under NDA. No confidential project, partner, model, task, score or output information is shared here.
Haoran “Tommy” Fang
Curiosity, with follow-through.
I’m studying Statistics and Data Science at UT Austin, with minors in Computer Science and Business. My work spans document intelligence, inference services, full-stack applications and AI evaluation.
The University of Texas at Austin
B.S. Statistics and Data Science GPA 3.93 · Graduating December 2026
Tools I work with
Python, TypeScript/JavaScript, SQL, Java, C++ PyTorch, TensorFlow, FAISS, Qdrant FastAPI, Flask, React, Next.js, PostgreSQL, Azure Docker, Terraform, Git and CI/CD
What’s next
Full-time software engineering and applied AI roles starting January 2027. Based in Austin and open to relocation. U.S. citizen; no sponsorship required.