Charles Jackson
I build software that puts a probability on the future — forecasting engines, agentic systems, full products — then grades every call against what actually happened.
Polypredict
See where the model disagrees with the market — and whether you can legally act on it.
Polypredict puts its own price-blind probability on every live Polymarket event, shows it against the market with an uncertainty range, and tells you — per bet, per jurisdiction — whether you can legally place it. Read-only by design. Often wrong, never advice.
The forecaster never sees the market price, so a disagreement is a real opinion — not an echo of the crowd.
Fractional Kelly over the joint outcome distribution, stress-tested for risk-of-ruin.
Your jurisdiction × the bet's category — a card tells you which, with a cited source.
Coherence checks surface real, fillable price inconsistencies across related markets.
- Python
- Forecasting
- Kelly sizing
- Legality engine
- Read-only
The model's own probability against the live market price — the gap is the edge, and the forecaster never saw the price.
Predictive
A calibrated probability on the decisions that move your company.
Predictive forecasts the outcomes you actually bet the business on — demand, churn, a launch, a market — as probabilities with honest uncertainty, then scores every call against what happened. No crystal ball: ranges, not promises. It runs the same forecasting kernel proven in public on Polymarket, pointed inward at your company.
Will you hit the number? A probability on the quarter — and the range around it.
Which accounts actually leave, scored against who really churned.
The odds a launch clears the bar, with the assumptions made explicit.
Rates, demand cycles, input costs — grounded in live FRED data.
- Calibration
- Next.js
- Supabase
- FRED data
- Invite-only
When the model says 70%, it should happen 70% of the time. Points on the diagonal are perfectly calibrated — that's the bar every forecast is graded against.
Digital Twins
Your own AI twin — a persona, a mandate, and a memory that persists.
digitaltwins.world is a platform for building an AI twin of yourself: a persona and mandate you define, memory that persists and consolidates what matters, a temperament that adapts to context, and twins that can negotiate and transact with each other on your behalf. Built end-to-end and shipped on Cloudflare Workers.
You set who it is and what it's allowed to do — its voice, its remit, its boundaries.
It remembers across conversations and consolidates what matters, so it gets sharper over time.
Regulated affect — equanimous in business, warm in relationship — so it never gets knocked off tone.
Yours can negotiate and settle with other people's twins on your behalf — an economy of agents.
- Digital twins
- Next.js
- Cloudflare Workers
- Persistent memory
- Voice
Built in Montréal.
I've spent the last two years on one question: can software put an honest probability on complex things — a market, a launch, a quarter — and prove it? Predictive is the engine that answer became. Polypredict is how it gets tested in public, price-blind against real markets, graded on what actually resolved.
I ship end-to-end — model, backend, product, design — because a forecast no one can act on is trivia. The craft is calibration: being right exactly as often as you say you'll be.
- Role
- Agentic systems engineer
- Base
- Montréal, QC
- Stack
- TypeScript · Python · ML
- Focus
- Forecasting & agents
- Now
- Building Predictive
Say hello.
Open to ambitious problems in forecasting, agents, and full-stack product.