Drop your bank, electricity, or fitness CSVs into a local folder and go to bed. By morning, open-weight TabPFN-3.5 foundation models forecast your next month and isolate every anomaly. Zero cloud bills. 100% private.
Multi-horizon walk-forward forecasting & seasonal outlier detection
Four planted demo storylines isolated with zero false positives
Fitness First gym raised monthly fee in June from $24.99 to $29.99.
City Power billed twice on the exact same date for the exact same amount.
Today vs median of 8 prior same weekdays. Flags exactly the 3 spike days.
Planted pipe leak in water.csv spiking to 425L on Oct 18 & 19.
Reproducible benchmark measured against GPT-4o-mini zero-shot prompting on the identical 30-day holdout window
| Attribute | Hare (Closed Hosted API: GPT-4o-mini) | Tortoise (Napper: TabPFN-3.5 Local) |
|---|---|---|
| Model Type | Proprietary Cloud API (gpt-4o-mini) | TabPFN-3.5 (Open-Weight Foundation Model) |
| Cost Per Run | ~$0.0123 / run ($4.49 / yr) | $0.00 (Zero runtime bill) |
| Data Privacy | ā Bank CSV transmitted over public internet | ā 100% Local (0 bytes transmitted) |
| Offline Capable | ā Fails without active internet connection | ā Works in airplane mode / air-gapped |
| Holdout Error (Daily Spend WMAPE) | 165.45% (Prompted zero-shot token regression) | 56.34% (3x better than GPT-4o-mini, 12x vs naive) |
| Monthly Spend WMAPE | 4.85% | 2.03% |
| Execution Latency | ~3.42s (Network API round-trip) | ~373s (Consumer AMD CPU "Slow & Steady") |
| Auditability & Portability | ā Closed black box; vendor can deprecate | ā Open weights, fully auditable & swappable |
End-to-end distributed telemetry monitoring offline agent tool calls with zero cloud fees
Multi-domain real-world and synthetic time-series benchmarked offline