Your herd already writes the data.
Now it answers back.
The Barnboard Data Platform turns every milking you log into governed metrics, open APIs, and a per-cow milk forecast — running on the same small computer as Barnboard itself, with nothing sent to anyone's cloud.
Try the live demo Read the documentation90-second silent walkthrough — the story, the PIN gate, and the live app with its forecast panel.
The problem
A dairy's records usually die in the notebook they were born in.
Numbers without meaning
Litres get written down twice a day, but "is the herd doing well?" still gets answered by gut feel. Averages computed by hand disagree with each other, and nobody can say where a figure came from.
Data locked in
Records trapped in one app — or one notebook — can't feed a vet's spreadsheet, a co-op's reports, or next year's loan application. Getting data out (and back in) shouldn't be a rewrite.
Problems found late
A cow that slips 20% over three days is telling you something — mastitis, heat, feed. By the time the drop is obvious in the monthly total, the cheap intervention window is gone.
What you get
🥇 A governed data pipeline
Every table flows through a bronze → silver → gold refinery with lineage recorded at each step. Metrics are defined once, governed always: the platform physically refuses to sum a ratio up a grain — the classic spreadsheet mistake that quietly corrupts averages.
📈 A forecast on the dashboard
A per-cow model (trained on your own records, standard-library Python, no GPU, no cloud) projects each cow's next 7 days. The dashboard shows the herd's expected week; a sustained drop against forecast flags the cow — in the demo it catches the seeded decliner with zero false alarms.
🔓 Open contracts, not lock-in
Everything the farm enters is available as CSV and a versioned JSON API with a schema endpoint any consumer can program against. Export, re-import, back up — the round-trip is tested, and the whole database is one portable file.
🔔 Alerts that earn trust
Withholding (DO-NOT-TANK), yield-drop early warning, and likely-heat detection — a cow on her 21-day cycle showing the tell-tale one-day milk dip gets flagged in time to breed her. Every alert needs two signals to agree, so quiet mornings stay quiet.
📊 A price outlook, honestly built
Next year's ₹/L estimate on the dashboard: 35 years of India's official milk price index give the direction, your own collection records give the level. Backtested before it was allowed on screen.
🌍 Benchmarks that fit your herd
Cited benchmark yields for the breeds Indian farms actually keep — Gir, Sahiwal, Red Sindhi, Murrah buffalo, HF and Jersey crosses — sourced from NDDB, ICAR and peer-reviewed studies, loaded right next to your own numbers for honest comparison.
Architecture
One machine, one stack, four layers. Data flows down; every layer is inspectable and every answer carries a receipt.
Feature matrix
What was asked for, what was actually built, and what's still ahead. A matrix that's all green is a matrix nobody believes — this one is honest.
| Requirement | What exists today | Status |
|---|---|---|
| Interoperable — data usable by other systems | Versioned JSON API (/api/v1/…), schema endpoint, CSV export and import (round-trip tested), one-file database with a rehearsed backup drill |
Built |
| Extensible — grow without engine changes | A new metric is one JSON definition; governance guard blocks unsafe rollups; extension path documented and demonstrated | Built |
| Governed pipeline with lineage | Bronze→silver→gold with per-step lineage, PII exclusion at ingest, re-runs proven bit-identical | Built |
| Free web data, properly researched | 16-source inventory with licenses; 3 datasets stored locally with sha256 checksums; gaps recorded openly (no open per-cow desi-breed data exists anywhere) | Built |
| Breed benchmarks incl. desi breeds | Six breeds with per-row citations (NDDB, ICAR, peer-reviewed), loaded into the gold layer | Built |
| Per-cow yield forecasting | 7-day damped-trend model; beats the naive baseline on a held-out week (MAE 1.25 vs 1.33 L) | Built (3 weeks of demo data — margins will grow with history) |
| Early-warning drop detection | Deviation-vs-forecast flags; catches the seeded decliner, zero false positives; half-logged days excluded so a missed PM entry never fakes an alarm | Built |
| Revenue projection | Month-end estimate from month-to-date rate; 8.1% holdout error on demo data | Built (demo-scale; needs multi-month history to trust for planning) |
| Inference served in the product | Predict API + dashboard forecast panel; model artifacts versioned and hot-reloadable with zero downtime | Built |
| Automated tests | 23-test suite over contracts, models, auth and rate-limiting — all passing | Built |
| Runs on bare metal, supervised | PM2 + systemd boot unit; kill-and-resurrect drill passes — all six services survive a power cycle | Built |
| Container deployment | Docker image builds and serves — login and dashboard verified running inside the container | Built |
| Milk-price outlook from market data | FAOSTAT India milk price index (35 years) gives the direction, your own sales give the level — a dashboard tile shows next year's ₹/L estimate, backtested (drift beats naive, 4.6% vs 6.9% error) | Built |
| Heat (estrus) detection | 21-day cycle + one-day milk-dip rules flag "likely heat" on the dashboard; zero false positives by construction (both signals must agree) | Built |
| Governed metrics API | Every governed metric served with its receipt — definition, owner, version, lineage, and the exact SQL (/api/v1/metrics) |
Built |
| Operations that run themselves | Nightly database backups (14-day retention, restore rehearsed) · nightly pipeline + model retraining · CI on every code change | Built |
| Encrypted access on the farm network | HTTPS on the shared lanes with secure cookies (self-signed until a domain brings a public certificate) | Built |
| Multi-farm / co-op analytics | Design drafted, deliberately parked by owner decision | Parked |
Roadmap
- The lactation curve takes over on its own — a curve-based forecaster already competes in every nightly retraining contest (it has beaten the trend model on a multi-month backtest); it ships automatically the night it wins outright on your real data.
- Narrated demo video — the hero walkthrough gains a voiceover.
- A public certificate — when a domain arrives, the self-signed HTTPS on the farm network upgrades to a browser-trusted one.
- Co-op view — parked by decision; the design is written and waiting.
Deployment notes
On-premises (the lane you're looking at)
The whole platform — app, pipeline, models, this site — runs on one modest machine under a process manager. Four pinned Python dependencies for the app; the data and model layer is standard-library only. Your data never leaves the building: models train and infer locally, and the reference datasets are fetched once and checksummed.
Cloud twin (when you want it)
The same codebase runs on a small VPS with no code changes — the
database is a single file you copy, and the demo seed makes a walkable synthetic
herd on first boot. A deploy/twin compose file exists from the app's
earlier package; the data platform's container lane ships when the Docker image
is verified (see the matrix — it's marked honestly).
See it with real motion
The demo is the actual product on a synthetic 12-cow herd — log a milking, watch the forecast panel, pull the API. Ask us for the access PIN if you don't have it.
Enter the live demo Documentation