I built my friend Chiboy Taryns — a TabPFN cashflow early-warning

I built Taryns for my friend Chiboy.

Chiboy freelances. Every Sunday he opens a spreadsheet, scrolls invoices, and guesses whether next week will be fine. Sometimes a client is late. Sometimes a retainer looks “in” when it’s still pending. He finds out when the week already feels thin.

Taryns turns his invoice CSV into an early warning:

1. Load invoices (date, client, amount, status, category, notes)
2. Aggregate into weekly cashflow
3. Run TabPFN to forecast next-week inflow, flag weird weeks, and estimate thin-week risk
4. Write a short Sunday brief he can act on — chase who, why this week looks off

It’s not another budgeting dashboard. It’s a gift: his spreadsheet finally talks back before he’s blindsided.

Live demo: https://taryns.vercel.app

Demo video; https://youtu.be/Qp9AECmmbDY?si=M20HNXLxGN0xa-sJ

Open it, keep the name as Chiboy, hit "Run sample forecast".

What TabPFN showed on his demo history:

• Next week forecast: ~$50
• Recent 4-week average: ~$925
• Thin-week risk: ~78%
• Weird week caught: 2025-12-15 — $0 collected while $2,100 from Oakline was still pending

I handed it to him live. He said:

"Wait… so it basically caught that I would've thought December was fine, but that Oakline money wasn't actually in yet? Yeah, send me this every Sunday. That's the part I always miss."

Repo: https://github.com/webski101/Taryns

Core pieces:

• src/tabpfn_engine.py — TabPFN forecast + residual anomaly detection
• src/data.py — CSV cleaning and weekly aggregation
• src/brief.py — Sunday brief from TabPFN outputs
• app.py — Streamlit app for full CSV upload locally
• web/ — Vercel demo of the Chiboy TabPFN run

How I Built It;

Open-source AI at the core is TabPFN (Prior Labs) — a tabular foundation model.

Why TabPFN and not a chat model for the hard part: Chiboy’s problem is a CSV. Dates, amounts, pending flags, client mix. TabPFN is built to forecast / classify / spot anomalies from historical tabular rows.

How it’s wired:

• Clean the invoice sheet and roll it up by week
• Local TabPFN v2 regressor predicts next-week inflow
• Compare recent weeks to what TabPFN expected → weird-week flags
• Thin-week risk from a TabPFN classifier signal on historical “thin” weeks
• A short brief layer explains those numbers in plain language (Ollama/Gemma when available; otherwise a deterministic brief from the same TabPFN outputs)

Remove TabPFN and Taryns dies. There’s no product left, just a spreadsheet viewer.

Stack: Python, Streamlit for the full local app, Next/static demo on Vercel for sharing the Chiboy run quickly.

Why Does Open Innovation Matter?

Freelance money data shouldn’t need a closed chatbot.

Open made this possible in ways a closed API wouldn’t:

• Local TabPFN — the prediction runs without a paid API; free to rerun every Sunday
• Data stays closer to him — invoice history doesn’t have to be the fuel for someone else’s chat product
• Swappable explanation layer — the brief can use local open weights via Ollama; the forecast stays TabPFN either way
• Inspectable pipeline — clean → weekly features → TabPFN → brief. No black-box “AI insights” glued on

Closed tools are great at sounding helpful. Open tabular prediction is better when the job is: learn from this sheet and warn me before I’m blindsided.

Not financial advice — an early-warning gift for one person.

Prize Categories

• Best Use of TabPFN — TabPFN is load-bearing: next-week inflow forecast + weird-week anomaly detection from historical invoice CSV data