gabe@kagan:~$ whoami
Gabe Kagan — I build production systems for fun. A few of them even make money.
gabe@kagan:~$ cat pitch.txt
By day: payments & fraud at a fintech. By night: a live trading bot moving real capital, an AI assistant living in my Meta glasses, and a RAG over 11 years of my own messages. Real systems, real users, real money.
gabe@kagan:~$ ls ~/live
kalshi-bot/visionclaw/personal-kb/vitals/local-llm-stack/
gabe@kagan:~$
Hi, I'm Gabe. 👋
I'm a product manager in payments and fraud — ten years deep in underwriting, risk decisioning, funds availability, and the unglamorous systems that quietly decide whether your transaction clears. I've done this at fintechs big and small.
Outside the day job I build things end to end — from the data layer to the deploy. The Kalshi bot below is 150K+ lines of Python running live on a VPS I manage. The personal knowledge base reads my own messages and answers questions about them. The Meta glasses see what I see. I lean heavily on modern AI tooling to move fast, and I'm at my best when I can own a problem from idea to production — whether I'm driving it myself or embedded with a team.
The Builds
Things I built that actually run. Click in for the gory details.
Kalshi Prediction Market Bot
Automated trading system for CFTC-regulated event contracts. EGARCH volatility model, Kelly-sized positions, calibrated probabilities — running live with real capital.
Vitals — quantitative AI coach on my own data
A personal health data layer: Oura, Renpho, and a decade of message behavioral signals in one local SQLite. A model-agnostic persona contract turns any LLM — Claude, GPT, Gemini, or a local 9B — into a coach that has to query before it speaks.
VisionClaw — JARVIS in my glasses
Real-time AI assistant on my Oakley Meta HSTN glasses. Streams what I see and hear to Gemini Live, which talks back and can take actions on my behalf. Built on the open-source VisionClaw + Meta Wearables DAT SDK.
Personal Knowledge Base
A RAG over 11 years of my own iMessages, WhatsApp, and email — 168,000 messages, queryable in plain English. Then a second pass that distills it all into a written portrait. 100% local if you want it.
Local LLM Stack
A cheap inference tier on a 24 GB MacBook — stdlib proxy, fallback chains, prompt cache, JSONL telemetry. The tier below Claude, not a replacement for it. Built so the mechanical 60% of LLM work costs me nothing.
Builder first. Consultant when it fits.
I take a few engagements a quarter, usually when the problem is weird enough to be interesting. Here's where I'm most useful.
AI-accelerated shipping
Your team has ideas but ships slowly. I help integrate AI coding tools into your dev cycle so you go from prototype to production in days, not months. Works best for startups and internal-tools teams that need velocity.
Fintech systems
Ten years across payments, fraud, and trading — production risk-decisioning at a proptech payments company, fraud systems in fintech, and a live Kalshi bot trading real capital. I can advise on product decisions where payments meet ML, or get hands-on with your risk and fraud work.
Stuck problems
Sometimes you don't need a team — you need someone to read the code, understand the constraint, and propose a fix. I'll spend a few hours, write it up, give you a path. No retainer.
Experience & Education
The formal timeline lives on LinkedIn — payments, fraud, and product, with noticeably fewer jokes.
View on LinkedIn