AI Agent
Stock statements, invoices, distributor reports, KYC packs — the documents that power your operations rarely arrive as clean data. Our extraction agent reads them the way a careful accountant would, validates against a typed schema, and writes the result straight into your systems.
Document understanding is more than OCR. The agent reads layouts, infers field meaning, validates types, and tags low-confidence cells for review — so what reaches your downstream system is reliable.
Pharma distributors send monthly stock statements as PDFs. Each one has a different layout. Our agent reads them all into one canonical schema — distributor metadata up top, dozens of SKU rows in the body, validated types throughout.
Identifying details extracted with the type and format the downstream system expects — not free-text strings.
Every SKU on the statement, parsed into a typed row — quantities as integers, values as floats, units normalised.
PDF arrives via the upload endpoint — UI, email-ingest, or a watched folder. The document gets a record and status of processing.
An RQ job is enqueued on a Redis-backed worker queue, so the API stays responsive while heavy extraction runs in the background.
A vision-language model runs against the typed Pydantic schema. Output is constrained to the shape your downstream code expects.
Field types are checked, totals are sanity-checked against line items, and low-confidence cells are flagged for review.
The structured result is saved alongside the source file, with an audit trail and a status of processed.
A side-by-side viewer lets a human compare the original PDF and the extracted fields, edit anything that needs correcting, and approve.
What you get back — a clean JSON document validated against the typed schema, ready to load into your warehouse, ERP, or analytics pipeline.
stock_statement_oct_2024.json
{
"distributor_name": "Aarogya Pharma Distributors Pvt. Ltd.",
"distributor_address_line": "Plot 14, MIDC Industrial Area, Phase II",
"distributor_city": "Pune",
"distributor_state": "Maharashtra",
"distributor_pincode": "411019",
"distributor_gstin": "27AABCA1234N1Z5",
"distributor_pan": "AABCA1234N",
"distributor_dl_no": "20B/MH-PUN/2019",
"distributor_fssai_no": "11522003000324",
"start_date": "Oct 2024",
"end_date": "Oct 2024",
"items": [
{
"product_description": "Paracetamol 500mg Tablets",
"packing": "STRIP",
"opening_qty": 1240,
"opening_value": 18600.00,
"purchase_qty": 5000,
"purchase_value": 75000.00,
"sale_qty": 4820,
"sale_value": 96400.00,
"closing_qty": 1420,
"closing_value": 21300.00
}
// ... 187 more rows
]
}A boring, reliable pipeline — which is exactly what you want when you're processing thousands of documents a month.
VLM Run for production extraction. Pluggable — we can swap in your preferred VLM provider or run it on-prem.
A typed API surface, a relational store for documents and extractions, and an RQ-based async worker queue.
Schema-as-code. The same types validate the model output, the API response, and your downstream consumers.
Stock statements, secondary sales reports, distributor returns.
Invoice and bill-of-lading capture into your ERP without manual entry.
Identity documents, address proofs, and licence packs into structured records.
Key clause extraction and structured summaries from long-form contracts.