Resume · Backend

Backend systems built around explicit failure modes.

Queues, crash isolation, GPU admission, cloud infrastructure, and durable local services—with the concurrency and recovery mechanics left visible.

Choose a focus

01

Experience

Nov 2025 – Present

CrunchAtlas

Software Engineer

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

  • 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.
  • Engineered the queue's crash-isolation so no job hangs non-terminal: each job runs in a spawned (not forked) supervisor that survives a segfault or OOM-kill and reaps the child's process group, while a heartbeat sweeper force-fails jobs whose supervisor dies or freezes (catching a frozen-but-alive one a PID check would miss). Chaos tests SIGKILL live supervisors and kill the database mid-job to prove it.
  • 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 in-house agent runtime the platform's security agents run 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 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.
  • Sole owner of AtlasCyber's AWS, end to end across GovCloud and commercial partitions: EC2 (with a reverse-proxy/tunnel fleet), S3, IAM, VPC networking, and CloudFront.
  • Rewrote the network-capture sensor from scratch in Rust, memory-safe workspace-wide (#![forbid(unsafe_code)]): a headless Linux daemon around a crash-safe SQLite-WAL upload queue that persists each capture before upload and survives a restart without dropping data, replacing the prior lossy in-memory queue.

Mar 2025 – Nov 2025

CrunchAtlas

Software Engineer, AtlasConnect

Sole developer and maintainer of an inherited codebase—Django REST, React/Vite, self-hosted LLM stack.

  • Built the document-ingestion pipeline for deal intake: multi-format extraction (PDF/DOCX/PPTX/DocSend) with selective parallel OCR on scanned pages, feeding concurrent LangChain chains under a managed token budget.
02

Selected projects

gloss

Python · SQLite/FTS5 · Ollama

  • Built a hybrid retrieval engine that fuses BM25 with a local-embedding vector channel via reciprocal-rank fusion to return a source document's verbatim, cited passages, tuned against a 266-case evaluation gated by a paired significance test (hit@5 0.94, hit@1 0.71).

Celery Fork-Safety Investigation

Python · Celery · GDB · LLDB · Docker

  • Root-caused a Celery prefork fork-safety bug to the register level: a Rust-backed vector store initialized before fork() left the child inheriting a lock whose owner didn't survive the fork—a permanent futex deadlock on Linux (traced in GDB) and a libdispatch semaphore crash on macOS (traced in LLDB). One cause, two OS-level failure modes.
03

Honors

  • 1st Place / Best Hack — Bitcamp 2024, UMD (largest US East Coast hackathon) — TermBook
  • Best Educational Hack & 2nd Overall — HackUMBC 2024 — Testif.AI