- Ran user interviews with non-technical professionals (tax, legal, medical) using AI agents on their real casework.
- Built a desktop agent that operates the computer through the terminal, with skills for Google Workspace CLI, Context7, Playwright CLI, OfficeCLI, and computer use via AppleScript and PowerShell.
- Created a GTM expert inside the product itself: sourced and screened 306 tax-resolution firms nationwide, generated 95 per-firm research dossiers, and drafted 63 fully personalized outreach emails.
Eddy Liang
Agent engineer — agents, the sandboxes they run in, and the evals that grade them.
Experience
- Led a small team to launch the company's flagship premium feature: code interpreter, command execution, and file upload infrastructure.
- Shipped a data analytics agent (36k+ queries/day) using chain-of-thought reasoning for multi-step analysis.
- Built a deep research agent (20k+ queries/day) that reads hundreds of sources in parallel to generate synthesized reports.
Collaborated with MIT Media Lab researchers. Built and shipped Almond, a Flutter app for collecting and labeling EMG and EEG signals mapped to facial expressions and body movements.
Projects
A minimal PyTorch re-implementation of Qwen vision-language models, built to understand multimodal transformers end to end: text+vision processing, dense and mixture-of-experts variants, streaming generation, and side-by-side validation against Hugging Face models.
A pipeline that turns an expert-corrected agent session into an anonymized, verifier-graded task environment. Regenerates case artifacts — receipts, filings, spreadsheets — with the underlying math intact but every identity and storyline replaced, runs a leak scan to certify the twin, and emits the task plus graders as a portable JSON world.
A QEMU/KVM desktop environment in Docker with a REST API and Python SDK for agents that need a real GUI. Screenshot, mouse, keyboard, and VM snapshot controls over QMP — disposable operating systems as programmable infrastructure.
Education
Skills
Python, TypeScript, React, PyTorch, Hugging Face Transformers, vLLM, FastAPI, Docker, Electron, Flutter, Tailwind CSS, QEMU/KVM