Skip to content
AI Games — software & AI studio, Sofia

PRODUCT · LIVE

Fatumu.com

Calibrated probability estimates for the questions that do not have clean answers. A three-model LLM ensemble (Claude, GPT-4.1, Gemini), an inverse-CI95 Bayesian average, and a polished PDF report — built by the studio on the same Cloudflare stack.

Type
B2B SaaS · web + iOS + Android
Status
Live
Ensemble
Claude Sonnet 4.6 + GPT-4.1 + Gemini 2.5 Pro
Output
Probability + 80 % CI + 95 % CI + drivers, in a PDF

What it is

A forecasting SaaS for founders, investors, traders and analysts. Ask a question with a verifiable outcome ("Will AAPL close above $250 on 30 June 2026?") and Fatumu runs it through a three-stage pipeline: ensemble inference, Bayesian calibration, narrative + Typst PDF render.

Why we built it

Most forecasting tools produce vague directional opinions. Fatumu produces calibrated probabilities — predictions that, over a sufficient sample, match observed reality. It is also the studio's proof-of-stack for ensemble LLM orchestration, calibrated outputs and report generation: everything we use here transfers directly to client work.

Stack

  • Next.js 16 on Cloudflare Workers (app, consumer, webhook)
  • Cloudflare D1, R2, KV and Queues
  • Cloudflare AI Gateway → OpenRouter (Claude Sonnet 4.6, GPT-4.1, Gemini 2.5 Pro)
  • Custom auth (Argon2id + HMAC pepper, PKCE OAuth, opaque sessions)
  • Stripe + Stripe Tax · Resend · Cloudflare Turnstile
  • Native Android (Kotlin / Compose) and iOS (Swift / SwiftUI) apps