CS @ Georgia Tech · Intelligence (AI/ML) and Modeling and Simulation
Building autonomous systems from real and simulated data.
I love building at the intersection where models meet the physical world, because I think computer science should complement our day-to-day lives, not replace them. Currently building Tycho, a deterministic verification layer for AI agents, and working on a perception model for Georgia Tech's HyTech Racing team.
|
A Windows file manager built around one thing: launching Claude Code from any folder in one click. Sessions open as tabs inside the app and live in named spaces you switch between, so a session waiting on a permission prompt marks its tab and its space instead of stalling unnoticed. That state is read from Claude Code's own hooks, not guessed from terminal output.
|
Electron, React and TypeScript, with real ConPTY terminals, a loopback MCP server that lets a session drive the app itself, bundled ripgrep search, and undo on every file operation.
Languages
Machine learning & AI
Web & app
Tools & infrastructure
| Project | What it is | Stack |
|---|---|---|
|
Checks whether an AI agent's tool calls actually did what the agent claimed. The verdict engine runs offline with no model in the loop, reasoning over git state, filesystem diffs, exit codes and database deltas. Open source for Claude Code, Cursor, Codex and OpenCode, and took it through a Y Combinator interview. | Python · LLM integration · PyPI/npm | |
|
Physics-informed NN on HTEM DB |
A year of mentored research on NREL's materials database. Built a ~9,500 sample by ~1,300 feature training set, then benchmarked 20 deep networks against a scikit-learn baseline for a 96.59% drop in MSE. Adding physics equations as learning constraints took the best architecture to an R² of 0.9932. | Python · PyTorch · scikit-learn · Jupyter |
|
Multimodal GNN for semiconductors |
Predicts band gap and formation energy straight from crystal structure across 17,000+ stable semiconductors from the Materials Project. R² of 0.81 on band gap and 0.98 on formation energy, with a custom PyTorch C++ extension that made CPU training ~2.5× faster. | Python · C++ · PyTorch Geometric · pymatgen |
Also contributed to: Pebble · Parky
Georgia Institute of Technology · B.S. Computer Science, 2029 · Atlanta, GA
LinkedIn ·
dmano6@gatech.edu