all systems in production

Works on my machine.
Works in prod.

I design, build, and run production software with AI. Multi-agent developer tools, full-stack SaaS, and quantitative trading systems, shipped solo by orchestrating Claude Code and Codex — from first spec to deploys, monitoring, and real users.

projects built & shipped
20+
lines of code in production
300k+
data points ingested by one system
12.5M+
live systems, real users & real money
24/7
~/status
mollyliveagent-collabprivate betaarvantoin productionpolymarketin productionlineup-impactin productionnhl-quantin developmentfutisin productionheartbeatin production

Projects

Sneak peeks from production

A selection of systems I've built and operate. Most run for real users or real money, so details are summarized — full case studies are on the way, and I'm happy to walk through any of them in depth.

molly — command desk

Molly — Market Intelligence

Real-time data analytics & decision platform

Live

Turns a live stream of market transactions into operational intelligence: 30-metric customer profiling, calibration analysis, forward-exposure forecasting, and statistical anomaly detection against institutional reference data. Cluster-robust statistics and empirical-Bayes models power an 11-desk operations dashboard used for daily decisions.

PythonFastAPINext.jsTypeScriptpandasscipyECharts
lines of code
82k
operational desks
11
customer-profile metrics
30
room · payments-refactor

Multi-Agent Dev Room

Real-time collaboration layer for coding agents

Private beta

A real-time collaboration layer for Claude Code, Cursor, Codex, and Gemini, built on MCP. Agents across machines join shared rooms to coordinate distributed development: code review, PR handoffs, and frontend–backend integration between sessions that otherwise can't talk to each other.

TypeScriptReactNode.jsMCPUpstash Redis
lines of TypeScript
101k
commits
330
agent clients supported
4
arvanto — dispatch

Arvanto

Configurable CRM/ERP for field-service organizations

In production

Multi-tenant CRM/ERP covering the full loop from service-job intake to PDF reporting and billing export into a Finnish accounting system. Claude-powered Excel import maps and classifies customer data automatically; row-level security isolates every organization across 84 database tables.

Next.jsTypeScriptReactSupabasePostgreSQLClaude API
lines of code
92k
database tables
84
test files
266
execution.log

Polymarket Trading Engine

Low-latency signal engine for prediction markets

In production

Streams Polymarket order books over WebSocket and compares them against reference prices in a sub-20ms decision loop. Fee-aware expected-value math, Kelly-criterion sizing, and six tiers of risk controls — position caps, exposure ceilings, daily circuit breakers — govern fully automated execution with real capital at stake.

PythonWebSocketasyncioPolymarket CLOBKelly criterion
scanner loop
<20ms
risk-control tiers
6
operation
24/7
lineup-impact — projection

Lineup Impact Engine

Forecasting lineup impact in pro & college basketball

In production

Quantifies how player absences shift expected game outcomes, then flags market prices that haven't caught up. Massey power ratings, Glicko-2 player ratings, and Box Plus/Minus feed a model validated walk-forward against professional market benchmarks — with live roster monitoring and Telegram alerts.

PythonSQLiteESPN APIGlicko-2Telegram API
teams modeled
360+
backtested matches
200+
markets
2
rating_model.py

NHL Quant

Ground-up forecasting stack for pro hockey

In development

A quantitative forecasting system for the NHL, architected spec-first before a line of model code: ingestion supervisors with schema validation, team-rating models with nightly retraining, an oracle-simulating backtester, chaos tests, and a Telegram advisory layer. Docker-first from day one; implementation is landing module by module.

PythonDuckDBDockerpandasTelegram API
architecture modules
12+
pipeline
ingest → model → backtest
deploys
Docker
futis — live feed

Futis

Lineup intelligence for Finnish football, tiers 1–4

In production

Tracks lineups, player ratings, and live match events across eight-plus Finnish football leagues where structured data barely exists. An autonomous scheduler polls official sources every 30–60 seconds, computes power ratings from 1.8M+ player appearances, and broadcasts updates to Telegram.

PythonFlaskSQLitepandasGoogle Sheets API
data points ingested
12.5M+
player appearances
1.8M+
leagues tracked
8+
heartbeat — feed status

Heartbeat Portal

Uptime monitoring for real-time data feeds

In production

Watches third-party real-time data feeds and alerts when delivery slows below its expected cadence — before stale data costs money. Five-second polling with intelligent batching, tiered anomaly thresholds with alert cooldowns, and a live status dashboard with historical gap analysis.

PythonFastAPINext.jsAPSchedulerDocker
polling cycle
5s
API endpoints
17
alert channels
Slack + webhooks

$ ls ~/more Also in the workshop: a real-time tennis-event alert engine, an AI job-search copilot with a Chrome extension, a market-data latency benchmark, and a dozen more — write-ups coming.

How I build

AI does the typing. I do the engineering.

My edge isn't writing every line by hand — it's orchestrating AI tools to produce systems that hold up in production, at a pace a solo builder couldn't otherwise touch.

  1. 01

    Spec before code

    Every project starts as a written spec: the problem, constraints, data model, and edge cases. AI agents are only as good as the brief — so I've become very good at briefs.

  2. 02

    Agentic build loop

    Claude Code and Codex do the heavy lifting while I direct: parallel agents for scaffolding, implementation, and review. I read every diff that ships. MCP servers wire the agents into real infrastructure — databases, browsers, deploy targets.

  3. 03

    Verify like it's prod — because it is

    Automated tests, adversarial AI code review, and hands-on verification against live-like data before anything deploys. Trading and billing systems don't forgive sloppy releases.

  4. 04

    Operate and iterate

    Shipping is the midpoint. Cron jobs, monitoring, logs, and error budgets keep systems honest — and every incident feeds back into the specs and the agent workflows.

Toolbox

What I work with

Chosen per project, not by habit — the constant is AI-first workflows and boring, reliable infrastructure.

AI tooling

Claude CodeOpenAI CodexClaude APIMCP serversMulti-agent workflows

Languages & frameworks

TypeScriptPythonNext.jsReactFastAPINode.jsTailwind CSS

Data & infra

PostgreSQLSupabaseRedisVercelCloudflareDockerGitHub Actions

Domains

SaaS & billingQuantitative trading & forecastingSports analyticsReal-time data pipelines

Contact

Building something with AI?

I'm open to AI engineer roles and collaborations — especially teams that want someone who has already shipped and operated AI-built systems, not just experimented with them.