Optimization is the easy half. People are the hard half. I work in the gap between the two.
Senior at Northeastern — business administration, finance concentration, environmental studies minor.
Ten systems vibe-coded in about eight weeks. 288 commits. Three running 24/7 in production.
AI writes the code; I hold the taste. Multi-agent orchestration, pgvector RAG, MCP tool servers, live-broker trading.
The work
benji-ai
A personal assistant I text on WhatsApp, vibe-coded end to end. Thirty-four tools, six of them locked behind a typed yes, and every turn measured. It can read its own source code — and it found bugs no test would have.
(NOT SO) INTELLIGENCE-OFFICE
An autonomous system that screens, sizes and places real orders through a live brokerage, unattended. It ran real money for 37 trading days and lost 5% while the market rose. The interesting part: the picks it chose to trade did far worse than the ones it passed over — and it caught its own ledger misreporting the damage.
Sales race track
A retro leaderboard that turns a sales floor's targets into a race on the office TV. Every rep is a car, a pace line shows who's on track, 100% sets off the checkered flag. Dependency-free, running all day on cheap TV hardware.
02 — Approach
Start with a conversation, not a spec
Before anything I sit with the people who'll actually use the thing and find the problem they really have — usually not the one written on the ticket. Then I build the smallest version that solves it and get it into their hands early, so I'm learning from real use instead of guessing. I'd rather ship one feature people reach for every day than ten they have to be talked into. I don't write the code line by line — I vibe-code: I direct the models and spend my own attention on what's worth building at all.
There's a short list of things I won't do. I won't build for problems nobody has. I won't oversell — I say plainly what a tool can't do, because that's what gets it adopted. And I won't call something working before it's verified: I built an AI trading strategy, ran the research, found no real edge, and rebuilt it from scratch. I'd rather retire my own work than defend it.