→ built because I had an old PC which was underutilized, and wanted to know exactly what happens between git push and a live URL.

Problem

Replicating a correct Vercel-style deploy pipeline meant solving a thundering-herd race on container wake-up, and enforcing a deployment state machine under concurrent BullMQ workers without corrupting state.

Solution

Hosted on my old PC, A Redis SET NX distributed mutex coordinates container wake-up so exactly one job fires regardless of concurrent requests, and a Postgres trigger enforces valid state transitions at the database layer instead of trusting application code to get the race condition right.

Result

Dual execution engine (AWS ECS Fargate for cloud, local Docker for bare-metal) behind one interface, sub-30s cold starts, and zero state-machine races under concurrent retry attempts.

→ built because grep and Ctrl+F stop working once a codebase spans six languages and you're the new person trying to understand it.

Problem

Cross-language dependency mapping needed to scale across large, multi-language codebases without the slow, ad hoc analysis most tools fall back on.

Solution

A multi-language AST parsing engine feeds a Neo4j graph database, paired with a RAG pipeline over that graph so architecture questions can be asked in plain English instead of grepped for.

Result

60% faster dependency resolution across 6+ language repos, sub-500ms query latency on 10K+ node graphs, adopted by 50+ beta developers for architectural debugging.

→ built for a hackathon deadline that didn't care whether SSE fan-out across five services was a good idea — it just needed to work in 24 hours.

Problem

Supply chain disruptions were typically discovered hours after the fact via email chains, with no automated path from detection to a ranked resolution.

Solution

Five independent Node.js microservices coordinated over server-sent events, with Gemini-powered agents scoring financial impact and generating ranked resolution strategies for human approval.

Result

Mean time to detect cut from hours to under 5 seconds, and a Supabase RLS-secured multi-tenant data layer that holds up under concurrent agent writes.