AI systems, grounded in the machinery underneath.

Mar 2025 – Present

Local-first, air-gapped AI security platform for critical infrastructure. Sole author of the systems below.

• Built the PCAP analysis agent, the platform's most-used investigation agent: an autonomous LLM works a packet capture end to end (protocol recon, DNS/HTTP/TLS, file carving, IOC extraction, timeline) into a cited report and triage verdict. It reasons only over the output of a real forensic toolchain it drives (tshark, Zeek, foremost), never raw packets, so every value it reports traces to tool evidence.
• Built PurpleHaze, an autonomous penetration-testing product: a local LLM drives Metasploit and real offensive tooling through a full multi-step engagement, holds session state across long runs, and produces compliance-grade OWASP and NIST reports, with no human in the loop.
• Built the in-house agent runtime every analysis agent runs on (LangChain/LangGraph): context compaction that counts real tokens and fails closed instead of overflowing, plus tool-output eviction, loop detection, and per-run container sandboxing (fresh, network-isolated, self-reaping).
• Built the Postgres-native job queue that replaced Celery for the platform's GPU, LLM, and sandbox jobs, so a burst of heavy AI work can't OOM an air-gapped box and any one job can be cancelled without killing its worker. It runs on atomic per-lane claiming (SELECT FOR UPDATE SKIP LOCKED), priority-FIFO ordering, dependency DAGs with cycle detection, and lock-free single-flight.
• Built the local LLM serving stack (llama.cpp behind llama-swap) with a byte-exact VRAM fit-proof: a Python model reproduces llama.cpp's KV-cache allocation to the exact byte across grouped-query attention, sliding-window layers, and per-K/V quantization, so a deployment's fit is proven in CI with no GPU or running server.
• Built GPU-memory admission control so concurrent LLM jobs can't overcommit one card: a per-pool Postgres advisory lock admits a job only when its declared context-token cost fits the pool's live budget, and a runtime READ COMMITTED requirement makes a claimer that waited on the lock re-price against just-committed usage instead of a stale snapshot.
• Built the agent that maps captured network activity to MITRE ATT&CK: capture-time validation rejects technique IDs the model hallucinates, and a measurable retrieval eval (Recall@k/MRR with regression floors) gates changes before any agent runs.

May 2024 – Aug 2024

• Built a retrieval-augmented chatbot for an NSF-funded Chesapeake Bay salinity project (SaltCast) at Horn Point Laboratory: FAISS retrieval over OpenAI embeddings feeding a streaming FastAPI/LangChain service; presented at the MIT Undergraduate Research Technology Conference.