AI agent infrastructure · Payments for agents · Security and developer tools
Computer Science student at Northeastern University, concentrating in AI, graduating May 2027. I build infrastructure for AI agents: how they pay for things, how they get authorized, how they drive a browser, how they understand code, and how a human can check what they actually did. My longer-term interest is AI safety and interpretability.
| Project | What it does | Tools |
|---|---|---|
| GatePoint | Security gateway for AI agents. Enforces policy on every tool call and audits it, with no changes to agent code. It uses a three-layer trust chain: Ed25519 challenge-response, short-lived JWTs, and CA-signed certificates. Each request carries a signature over RFC 8785 canonical JSON. The policy engine supports 18 JSONPath operators, and policies can be generated by GPT-4 and validated with Zod. 2nd place at the Husky Startup Challenge. | TypeScript, Express, PostgreSQL, Prisma, React |
| Stepglass | Self-hosted tool that records what an AI agent did and verifies what actually changed in the systems it touched. It labels each change with before-and-after evidence and checks the result against policy. Built in public under Apache-2.0. | TypeScript, Node.js |
| openBrowser | Rust CLI that drives one persistent Chrome tab. Each command runs as a short-lived process, while a local daemon keeps the session alive. Gives structured JSON output and is tested against a real Chrome. | Rust, Chrome DevTools Protocol |
| Neocortex | Turns a GitHub repository into a knowledge base you can query. It parses repos into AST entity graphs at about 1,000 entities per second and returns ranked, token-budgeted context to an LLM in under 500 ms. My parts were embeddings and semantic search, program-slicing compression, and the extraction commands. | TypeScript, tree-sitter, OpenAI, Voyage AI |
| ImageDoctor | Full-reference image quality assessment with a fine-tuned ViT-Base on KADID-10k. Uses about 70% less GPU memory through selective layer freezing and gradient accumulation. | Python, PyTorch, timm |
| Loan Default Prediction | Credit-risk model on 1.35M Lending Club loans, using only application-time features. Reaches 0.72 AUC-ROC with XGBoost, with SHAP interpretability and error analysis by loan grade. | Python, XGBoost, DuckDB, SHAP |
More work: Marketing A/B Test Analysis · Memora · FUSE filesystem in C · open-instinct
| Organization | Role | Dates |
|---|---|---|
| Allowance (Y Combinator) | AI Product Engineering Co-op | Jul 2026–Dec 2026 |
| GatePoint | Founder & Lead Engineer | Sep 2025–Feb 2026 |
| Rezolve.ai | Software Development Engineering Intern | Jul 2025–Dec 2025 |
| Zolve | Data Engineering Intern | Jun 2024–Sep 2024 |
My work spans payment infrastructure for autonomous purchasing, issuing one-time virtual cards with approval controls and spending limits. It also covers a native Android app and MCP evaluation tooling at Allowance, cryptographic security for agent tool calls, real-time LLM sentiment pipelines with Redis caching that halved API costs, and a KYC verification pipeline that processed 15,000+ documents.
- HackBeanPot 2025, 1st place: Memora, an Alzheimer's remediation app that generates memory quizzes from tagged photos. Built in under 36 hours.
- Husky Startup Challenge, 2nd place: GatePoint.
- Languages: TypeScript, Python, Kotlin, Rust, C, SQL, Java
- AI and agents: OpenAI and Anthropic APIs, MCP, Browser Use, PyTorch, scikit-learn, XGBoost, Hugging Face
- Applications and infrastructure: React, Next.js, Node.js, Express, Jetpack Compose, PostgreSQL, Prisma, Redis, Supabase, Docker