CalibreThe movement, part by part.
What reads your invoices, what checks them and what decides. Calibre is document AI built in-house for Indian accounts payable: our own model, computer vision we tune ourselves, a rules engine that always runs, and confidence calibrated to your outcomes.
Specification
- Reader
- Our own document model: an open-source model fine-tuned for Indian invoices and hosted by usPinned by version; every value records which model read it.
- Grounding
- Every value must point to words on the pageA value the model cannot locate is never cleared on its own.
- Rules engine
- Runs beside the model on every invoiceSo the model is never the only reader.
- Second read
- Uncertain fields are read again from a different view of the page
- Proof
- Four tiers: form, arithmetic, GST and identity, policy
- Calibration
- Confidence fitted to verified outcomesAn invoice clears on its own only when the evidence supports it.
- Learning
- Vendor memory, rules, recalibration and gated model releases
- Languages
- Invoices from across India, including regional-language invoicesThe interface and reports are in English.
- Compute
- Private GPU infrastructure we operateNo third-party AI API, enforced in code.
- Isolation
- Tenant-scoped data, row-level security and a key per customer
Reading the page before reading the invoice.
Before a model sees a word, the page is turned upright, its layout is found and every word is placed with its position. Native PDF text is used exactly; text added by other software is labelled, not trusted.
- i
Orientation
Sideways and upside-down scans are turned upright.
- ii
Layout
Regions, tables and reading order, with every word’s position.
- iii
Long invoices
Read in parts, with every row counted against the printed total.
- iv
Splitting
One PDF holding several invoices is split at each signed e-invoice.
How a new reader earns its place.
Calibre’s reader is fine-tuned for Indian invoices, and every new version must prove itself before it reads a single customer invoice.
- 1
Pool
Only from customers who opt in. Off by default.
- 2
Pseudonymise
GSTINs, PANs, names and bank details are replaced before anything is pooled.
- 3
Train
Adapters trained on the pooled set, then sealed into a versioned bundle.
- 4
Gate
Must beat the current reader on vendors it has never seen, pass a leakage test and hold its accuracy with no history.
Corrections close the loop. Gates keep it safe.
Every correction enters the circuit. Nothing changes what Calibre does until it has passed its gate: a background trial for rules, and held-out vendors and a leakage test for models.
- Every correction is kept as evidence, with where it sits on the page.
- Verified invoices teach Calibre each vendor’s conventions.
- Five corrections that agree become a rule, proven in the background first.
- Confidence is recalibrated as verified outcomes accumulate.
- Each new reader is trained on pooled, pseudonymised data from customers who opt in.
Nothing inside calls an outside AI.
Your documents are read by models we host, on private GPU infrastructure we operate. A dependency check in our build blocks any library that would send them elsewhere.
- Every customer isolated at every layer, with its own encryption key.
- Services pass pointers, never documents; no invoice content in logs.
- Every upload scanned for malware, and refused if the scan fails.
- Single sign-on with OpenID Connect.
- Built to keep your data in India.
Questions finance teams ask
Is Calibre built on ChatGPT or another public AI?
No. Calibre’s reader is our own document model: an open-source model fine-tuned for Indian invoices and run on infrastructure we operate. Your invoices are never sent to a third-party AI service.
How does Calibre avoid AI mistakes?
Every value the model reads must point to words on the page, a rules engine reads beside it, uncertain fields get a second read, and four tiers of checks run on every invoice. An invoice clears on its own only when calibrated confidence supports it.
How does the model improve?
Corrections become vendor memory and rules straight away. Confidence is recalibrated as verified outcomes accumulate, and each new reader is trained on pooled, pseudonymised data from customers who opt in. It ships only if it beats the reader before it.
See it on your own invoices.
Send us a month of invoices. We replay them against your ERP’s records and show you, vendor by vendor, what Calibre would have cleared, and why the rest would have gone to a person.