# David Gao

> ML Researcher & Software Engineer at Purdue University, B.S. Computer Science & Data Science, Expected May 2028, GPA 3.72. This is a structured, LLM-friendly version of davidgao.com, curated for agents reading on a user's behalf.

i like making cool stuff and i also like piano and games

## Projects

- [DeltaVision](https://github.com/ddavidgao/deltavision): Observation middleware for browser-based GUI agents. Open-source middleware that detects meaningful screen changes between AI agent actions so the agent can skip redundant screenshots. MIT-licensed, with adapters for Anthropic Computer Use, OpenAI CUA, Browser Use, Skyvern, and Stagehand. _Tech:_ Python, OpenCV, FastAPI.
- [DeltaVision-OS](https://github.com/ddavidgao/deltavision-os): OS-level delta-first agent framework (V2). Extends DeltaVision beyond the browser to desktop apps and OSWorld VMs. Same CV cascade, but observation source is mss-based screen capture and the action space adds drag, double-click, right-click, and multi-key hotkeys. Runs against OpenAI-compatible VLM endpoints including llama.cpp server. _Tech:_ Python, mss, pyautogui, OpenCV, Ollama, llama.cpp, OSWorld.
- [DG Attention](https://github.com/ddavidgao/dg-attention): Experimental attention mechanism for OpenAI's Parameter Golf. Alternative attention mechanism where deeper layers transmit only inter-token changes via a parameter-free depth schedule. Zero added parameters, and within 0.003 BPB of standard attention. Written up as a negative result: DG measurably reorganizes the value stream (variance ratio 0.36 vs 1.38 at deep layers) without that translating into a quality gain. Named one of OpenAI's 3 favorite non-record submissions (out of 2,000+) in their Parameter Golf retrospective; merged to the official openai/parameter-golf repo. _Tech:_ Python, PyTorch, CUDA, Distributed Training, 8xH100.
- [Slinkt](https://slinkt.app): Large file transfer without the paywall. File transfer SaaS for large uploads using chunked multipart transfers to Cloudflare R2. Launched on Hacker News, with Stripe billing, password-protected downloads, custom slugs, and IP-based rate limiting. _Tech:_ Next.js, TypeScript, tRPC, Cloudflare R2, Stripe, Prisma.
- [CardboardAI](https://cardboardai.dev): AI-native storage management. Multi-tenant platform automating storage facility operations. Custom LangGraph agent handles tenant onboarding, invoicing, and payment collection. Dual-portal architecture serving operators and tenants. _Tech:_ Next.js, TypeScript, AWS RDS, Prisma, LangGraph.
- IoT Locker System: Commercial hardware + software. Smart locker system built on Supabase Realtime and Firebase, deployed across 40+ units with over $100,000 in revenue. A later migration to a Flask + Chromium kiosk made the Raspberry Pi stack lighter and faster to boot. _Tech:_ Flask, Supabase, Firebase, Raspberry Pi, Chromium Kiosk.
- [Resume Builder MCP](https://github.com/ddavidgao/resume-mcp): AI-powered resume tailoring via MCP. Full-stack MCP server for automated resume optimization. Maintains a SQLite profile database, scrapes job postings, and generates tailored LaTeX resumes with relevance-scored experience selection. Deterministic ATS analysis catches keyword gaps while the connected LLM handles semantic rewrites. No extra API keys needed. _Tech:_ Python, SQLite, LaTeX, MCP Protocol, openpyxl, trafilatura.
- [Purdue Dining](https://github.com/ddavidgao/purdue-dining-mcp): AI dining assistant for Purdue, works on ChatGPT & Claude. Say 'I'm hungry' and get real food recommendations from live Purdue dining menus. Deployed as a remote MCP server on Railway, serving both Claude (via MCP connector) and ChatGPT (via GPT Store with REST Actions). Pulls real-time menus and hours from Purdue's HFS API. Never guesses or fabricates menu items. Zero user data stored server-side; preferences live in ChatGPT memory or Claude's project context. _Tech:_ Python, MCP Protocol, Railway, Starlette, Purdue HFS API.
- [RLM](https://github.com/ddavidgao/RLM_TEST): Grounded AI through code execution. Agentic RAG system that forces LLMs to search documents via Python REPL instead of hallucinating. Multi-model orchestration with sandboxed code execution and evidence-based answer grading. _Tech:_ Python, Ollama, DeepSeek, Qwen, Matplotlib.

## Experience

- **Software Engineer Intern**, TalentPluto (Jul 2026 - Present). YC-backed AI recruiting platform (San Francisco). Own the candidate side end to end: the consumer onboarding funnel, the AI visibility score that shows candidates how AI models describe them, and the backend underneath both. Cut the pipeline history query from 16.1s to 0.58s with targeted indexes, and the signup-to-dashboard path from 13.7s to 9.2s.
- **Founding Engineer**, Fleet (Jul 2026 - Present). macOS desktop app (Tauri 2, Rust core, React 19) that runs fleets of parallel Claude Code agents. Built worker admission and lifecycle recovery, the Captain chat surface that survives restarts and 140MB transcripts, and git worktree isolation so agents can work the same repository without colliding.
- **Software Engineer Intern**, Pipelines (Apr 2026 - Jul 2026). Pre-deployment agent simulation platform backed by Sierra Ventures (San Francisco). Designed a synthetic data pipeline that uses real records as context to generate schema-valid test profiles. Built forward-deployed eval harnesses for RentAHuman (YC P26) spanning trace collection, world-state validation, and judge pipelines; deterministic payment and dispatch gates are projected to cut unsafe actions by roughly 85%.
- **Developer**, BoilerMake (Dev Team) (Apr 2026 - Present). Built and shipped full-stack features across BoilerMake's application platform, spanning frontend, backend services, and data infrastructure that thousands of applicants will use.
- **Software Engineer Intern**, Quture (Mar 2025 - May 2025). Pre-seed AI clothing marketplace. Built virtual try-on features using the OpenAI SDK to overlay clothing onto user photos; cut product return rates by 25%. Responded to a compromised Stripe credential incident, shipped emergency patches, and hardened the authentication flow.
- **ML Research Assistant**, Washington University in St. Louis (May 2024 - Aug 2024). Implemented an evaluation pipeline for self-supervised medical image reconstruction using PyTorch and NumPy. Ran diffusion bridge experiments across subsampling levels on WashU's GPU clusters, contributing to an IEEE ISBI 2025 publication.

## Skills

- **Languages:** Python, TypeScript, JavaScript, Rust, SQL, Java
- **Frameworks:** FastAPI, Next.js, React.js, Tauri, Flask, PyTorch, TensorFlow, Celery, tRPC
- **Infrastructure:** GCP, AWS (EC2, S3, RDS), Docker, Redis, Firebase, Supabase, Cloudflare R2, Vercel
- **Tools:** PostgreSQL, Prisma, LiteLLM, FastMCP, Stripe API, Git, Raspberry Pi

## Contact

- [GitHub](https://github.com/ddavidgao)
- [LinkedIn](https://www.linkedin.com/in/david-gao-322837355/)
- [X](https://x.com/ddavidgaoo)
- [Instagram](https://www.instagram.com/dvaidgao/)

## Optional

- [Human-readable AI page](/ai): same info, rendered as a page
- [Main site](/): full portfolio with styling