Open to AI roles · Remote or Paris · English

PonyRivers

AI Workflow Engineer
Solo Founder

I turn messy real-world inputs into reliable systems. In 13 months I went from my first line of Python to running a production SaaS on my own — API, database, payments, community, and a prediction engine that two AI assistants operate together.

~/pony — build.log

      
0Public repos shipped
0Live, deployed apps
0Production SaaS · solo
0Model ATS · 92-game calibration
01 · The story

Zero to production
in 13 months.

I'm an American living in Paris. In October 2025 I started my AI journey with Harvard's CS50 and CS50 AI — and I decided early that I'd learn by shipping, not by watching.

My first tools came from real workflows: agency itineraries that needed to become calendars, bank statements that needed to become tax buckets, portfolios that needed to be assembled from scattered links. Each one had to work in 30 seconds with sample data, so anyone could try it.

Then I picked a hard, measurable problem — pricing NFL games — and kept building until it became a business: Always Collect, a live SaaS with a FastAPI backend, Postgres, Stripe billing, tier-gated content and a Discord community. I built it alone.

Today I run APIE, my own learning engine for team ratings and player projections, and LIFE OS, a command center that has ChatGPT and Claude work from the same verified data, tracked by run ID.

"I don't stop at the demo. I ship, measure, and fix."

What I'm looking for: a team building AI products where I can own workflows end-to-end — data in, model or LLM in the middle, a clear, trustworthy output at the end.

02 · Flagship systems

Systems, not scripts.

Three connected builds that run every week of the NFL season. Click any box in the diagram to see what it does — or run a scenario to watch data move through it.

A · Production SaaS
Always Collect Platform

NFL betting-intelligence platform. Built, deployed and operated solo — soft launch June 21, 2026; live season picks from Week 1.

alwayscollect.com ↗
FastAPIAPI on Railway
PostgresSupabase · row-level security
Stripe3 paid tiers · webhooks
AutoPay → account → access

Architecture

B · Decision engine
NFL POWERBOOK

Odds in → probabilities → expected value → stake size → logged, settled and graded against the closing line. Treats betting as a pricing + risk + process problem.

Case study repo ↗
54.3%ATS · 92-game calibration
0.2284Brier score (lower = better)
52.4%Break-even at −110
264Players rated · 32 teams

Try the engine

illustrative math · production weights private

Pipeline

Honesty note: 92 games is a small sample. The 2027 off-season plan is a backtest on 5–6 past seasons (1,500+ games) and a logistic-regression rating → win-chance layer. Winner accuracy is ignored on purpose — only results against the spread count.

C · Learning model
APIE — Adaptive Predictive Intelligence Engine

My own model that learns from play-by-play, stats and schemes to produce weekly team power ratings and top-5 player projections per team — then gets graded against the POWERBOOK every week.

How APIE runs

03 · AI orchestration

LIFE OS // command center

The front door to everything: one screen that opens the live POWERBOOK, hands work to ChatGPT (V3 intel) and Claude (APIE), and proves both are reading the same verified data — every file stamped with its week and run ID. Below is a working mini version. Press the buttons.

Snapshot with demo data — the real command center connects to my Google Sheets, Drive exports and two AI assistants.

04 · Apps lab

11 apps. All live.
Try them right here.

Every app below runs in this page — I ported the core Python logic from each repo to the browser, so there's no sign-up and no cold start. (Where the full app needs live web search or a microphone, the demo runs the same logic on sample input.) Each lab also has the architecture, how it works, the real app, the code and the demo video.

05 · Toolkit

What I build with

06 · Résumé

The one-page version.

Pony Rivers

AI Workflow Engineer · Python · Automation · RAG
Paris, France · Open to remote
mrponyrivers@gmail.com
ponyrivers.netlify.app
github.com/mrponyrivers
alwayscollect.com
youtube.com/@PonyRivers-x7i
Summary

Self-taught AI workflow engineer who ships. Since October 2025 I've published 13 repos and 11 deployed apps, and built and operate a production SaaS alone — from FastAPI and Postgres to Stripe billing and automated member access. I specialise in turning messy inputs (statements, itineraries, PDFs, odds feeds) into reliable, auditable outputs, and in orchestrating multiple AI assistants around one source of truth.

Experience
2026 — nowFounder & Solo Engineer — Always CollectNFL betting-intelligence SaaS · Paris
  • Built and deployed the full stack alone: FastAPI on Railway, Supabase Postgres (row-level security), Netlify multi-page site, Streamlit member dashboard, Discord community.
  • Automated onboarding: Stripe webhook creates the user, writes their tier, and emails a magic login link — no manual steps.
  • Built an in-house CMS (Postgres table + admin API + tier-gated reader) and seeded 37 articles.
  • Designed the NFL POWERBOOK pricing engine: odds normalisation, implied/no-vig probability, EV, fractional Kelly, bet log, settlement and CLV tracking. Calibration: 54.3% ATS over 92 games, Brier 0.2284.
  • Building APIE, a learning model for weekly team ratings and player projections, graded weekly against the POWERBOOK.
  • Built LIFE OS, a command center that routes work between ChatGPT and Claude with week + run-ID provenance on every file.
  • Ran self-audits that caught and fixed real data bugs (wrong-position stats, missing neutral-site flags, team-code mismatches).
Oct 2025 — nowAI Workflow Engineer — Independent portfolio13 public repos · 11 live Streamlit apps · demo videos on YouTube
  • RAG PDF Assistant — local embeddings retrieval with similarity scores and page citations; cached indexing by file hash.
  • Tax Buckets — statement parser + rules engine with confidence scores, review workflow and QuickBooks-style export.
  • Itinerary Calendar Builder — free text → WORK/TRAVEL/HOLD .ics calendars with RunID tagging for clean bulk-delete.
  • French Pronunciation Coach — speech-to-text + text-to-speech with the OpenAI Audio API and a similarity scorer.
  • Betting-math suite: Odds/EV calculator (CLI + web, tested), Bet Decision Trainer with backtests, Line Shopping + CLV tracker.
Education

Harvard CS50x — Introduction to Computer Science (edX) · in progress
Harvard CS50 AI — Introduction to AI with Python (edX) · in progress

Core skills
PythonFastAPIStreamlitpandasSQL / PostgresSupabaseREST APIsWebhooksStripeGit / GitHubRailwayNetlifyHTML / CSS / JS
AI & data
RAGEmbeddingssentence-transformersOpenAI APISpeech-to-textPrompt workflowsMulti-assistant orchestrationModel calibrationBrier scoreBacktesting
Workflow
ETLParsingData cleaningCSV / JSON / XLSX / ICSGoogle Sheets automationPython video pipelines (PIL + ffmpeg)
Shipping habits
  • Every app demos in ~30 seconds with sample data
  • Exports for every output
  • Run IDs and audit trails
  • Honest metrics over flattering ones
Languages

English — native
French — learning (built two apps to practise)

07 · Let's talk

Let's build
something real.

mrponyrivers@gmail.com · Paris, France · Remote-ready