Conjuncts, matras, half-characters
Devanagari fuses consonants into new shapes, hangs vowels on four sides of a character, and joins whole words under a headline stroke. Segmenters built for Latin cut them apart.
क् + ष = क्ष
Your OCR can't read this.
श्री बालाजी किराना भंडार — कुल योग ₹633.50
Foreign OCR is trained on Latin typography and clean European invoices. Point it at a Devanagari land record or a faded carbon copy and it returns truncated characters in the wrong order. Pluton reads the page the way a person does — layout, tables, handwriting and script together — and returns structured data you can write straight into a system.
Send us a document See how it works In development · Devanagari first
Built in Chhattisgarh Read in every Indian language
HEADING · 0.97
श्री बालाजी किराना भंडार
बागबहरा, महासमुंद, छत्तीसगढ़ · GSTIN 22AABCU9603R1ZX
TABLE · 0.88
| चाय पत्ती — 500 ग्राम | 245.00 |
| सरसों तेल — 1 लीटर | 178.50 |
| आटा — 5 किलो | 210.00 |
| कुल योग | 633.50 |
HANDWRITING · 0.61
प्राप्तकर्ता के हस्ताक्षर — 04/09/2026
Extracted fields
6 fields · 1 needs review · 1 flagged
merchant_name
श्री बालाजी किराना भंडार 0.97
gstin
22AABCU9603R1ZX 0.95
line_items
3 rows parsed 0.88
total_amount
₹ 633.50 0.99
date
2026-09-04 0.93
signature_block
प्राप्तकर्ता के हस्ताक्षर 0.61
Reads
GST invoicesखसरा / खतौनीe-Way billsThermal receiptsChequesPrescriptionsStamp papersGST invoicesखसरा / खतौनीe-Way billsThermal receiptsChequesPrescriptionsStamp papers
Measured, not asserted
Numbers, and the documents they came from.
Most accuracy figures in this category are measured on clean benchmark sets. Ours are not. Every number below comes with the conditions attached.
99.4%
Printed Hindi accuracy
Complex Devanagari ligatures
86.8%
Handwritten Indic accuracy
Field notes and revenue signatures
100%
Table layout fidelity
Multi-column headers, row spans
<0.4s
End-to-end latency
Standard 300 DPI page
Evaluated on an internal benchmark of 10,000+ authentic Indian documents — tax invoices, handwritten land records and degraded carbon copies. This is our own eval set, not a public one; we publish the conditions alongside the figures so they can be judged properly.
Built for the hard part
Everything that breaks when the page isn't English.
Generic OCR drops below usable accuracy on Devanagari, and falls apart entirely on a handwritten ledger photographed under a tube light. That failure is the whole product.
Tables that survive
Multi-column headers, row spans and line-item grids come back as structured data — not a run-on paragraph where the numbers used to be.
Confidence on every field
Each value carries a score. Below 0.70 it is flagged for a person instead of being passed downstream wearing the same face as a certainty.
CLAHE restoration for the pages nobody designed for
Contrast Limited Adaptive Histogram Equalization, contour detection and affine homography deskewing run before anything is read — recovering glare, deep shadow, skew and faded toner. Preprocessing is the single largest accuracy lever we have measured, ahead of swapping the recognition model.
Nothing has to leave the building
Hybrid edge and cloud. On-device inference for offline scanning; no raw image retained after processing. Aligned to the DPDP Act, 2023.
Where it goes to work
The documents Indian institutions actually run on.
Not "invoices and receipts" in the abstract. The specific forms, in the specific scripts, that currently get retyped by hand.
01 / BANKING
Fintech & KYC
PAN, masked Aadhaar, passbooks, cheques and ITR forms extracted and validated, with the fields that matter scored individually rather than as one document-level number.
02 / GOVERNMENT
Revenue & land records
खसरा and खतौनी land records, ration cards, Gram Panchayat filings and judicial notices — the documents where table structure is the whole point and where nothing may leave the district office.
03 / LOGISTICS
Supply chain
e-Way bills, multi-item delivery challans and transport receipts parsed at speed, with line items reconciled instead of returned as a paragraph.
04 / LEGAL
Contracts & filings
Agreements, bilingual stamp papers and notarised affidavits turned into searchable text and vectors — bilingual pages handled as bilingual, not as English with noise in it.
05 / HEALTHCARE
Claims & records
Handwritten prescriptions, diagnostic reports and discharge summaries. The hardest input we handle, and the one where the confidence score earns its place.
06 / DEVELOPERS
REST API & SDKs
Drop-in SDKs for Python, Node and Kotlin with batch processing and webhooks. Layout preservation and target languages are request parameters, not a separate product.
One endpoint
Send a page. Get a schema.
REST with SDKs for Python, Node and Kotlin. Layout preservation and target languages are request parameters. Every response carries coordinates and per-field confidence, so you decide what to trust.
# Pluton Document Intelligence — Indic VLM import requests url = "https://api.plutontechnology.com/v1/ocr/indic" headers = {"Authorization": f"Bearer {API_KEY}"} with open("land_record_khasra.jpg", "rb") as doc: r = requests.post( url, headers=headers, files={"document": doc}, data={ "preserve_layout": True, "target_languages": ["hi", "en"], }, ) data = r.json() data["confidence_score"] → 0.994 data["structured_data"] → { … }
Illustrative request. The API is in development — endpoint paths and field names may change before general availability.
Stage, stated plainly
Where this stands today.
One product in users' hands, one still being built, and a door that is open. We would rather draw that line ourselves than let a demo imply it is somewhere else.
In development
Document Intelligence API
The Indic vision-language model at the core of our platform, exposed as one REST endpoint. Interfaces are still moving; endpoint paths and field names may change before general availability.
Open now
Institutional pilots
Banks, government departments and logistics operators with a real document backlog. Send us the pages your current tool returns empty and we will run them.
Questions we actually get
The things worth asking first.
Including the ones with an inconvenient answer. If yours is not here, write to us and it probably will be.
Ask us directlyWhat does Pluton actually do?
We build document intelligence for Indian languages. A vision-language model reads a page the way a person does — layout, tables, handwriting and script together, in one pass — and returns structured data with a confidence score on every field, rather than a stream of loose text.
How is this different from Google Document AI, AWS Textract or Azure?
Those are strong, general systems whose priority is not Devanagari. Each is also tied to its own cloud, which rules them out for the records that are legally not allowed to leave a district office. We are narrow on purpose: Indic scripts first, layout and tables as the primary output, and a deployment model where nothing has to leave the building.
Which languages and scripts do you handle?
Devanagari first — Hindi and Marathi — alongside English, and the mixed bilingual pages that most real Indian documents actually are. The model is trained on conjuncts, matras and half-characters rather than adapted to them after the fact. Further Indic scripts follow the same method.
Are your accuracy numbers independently verified?
No, and we would rather say so than imply otherwise. The figures on this site come from our own benchmark of 10,000+ authentic Indian documents — tax invoices, handwritten land records and degraded carbon copies — not from a public leaderboard or a third-party audit. We publish the conditions next to every number so they can be judged properly, and we would rather be measured on your documents than on ours.
Do you store the documents we send you?
No raw image is retained after processing. Captures are processed and purged, history can be cleared in one tap in the app, and the platform is aligned to the DPDP Act, 2023. Where a document may not leave the premises at all, inference runs on-device instead.
Can it run offline, or inside our own infrastructure?
Yes. The architecture is hybrid edge and cloud. On-device inference handles offline capture and the cases where data cannot leave the building; the cloud path is for throughput. The same engine backs both, so accuracy does not depend on which one you choose.
What is available today, and what is not?
The REST API is in active development and not generally available — the sample request on this page is illustrative, and its field names may still change. Institutional pilots are open today.
Who is behind Pluton, and where are you based?
Pluton Technologies (OPC) Private Limited, founded in 2024 and registered in Mahasamund, Chhattisgarh. We are early and small, and the fastest way in is to write to the founder directly at support@plutontechnology.com.
How do we start a pilot?
Send us a document — ideally the worst one you have. The faded carbon copy, the handwritten खसरा, the page your current tool quietly returns as empty. We run it, show you the output field by field with confidence scores attached, and you decide from there. No procurement process to start a conversation.