Drishtiदृष्टि
- Status
- Generally available
- Field accuracy at review threshold
- 99.2%
- Documents / month, largest deployment
- 340k
- Median end-to-end per document
- 14s
Document AI for the languages everyone else skips.
The part everyone underestimates
Extraction accuracy is not the hard problem. Knowing which extractions you can trust is. A model that is 94% accurate and silent about which 6% it got wrong forces a human to check all of it, which means you have automated nothing. Drishti reports confidence per field, not per document, and gates on it.
Scripts, not just languages
Devanagari conjuncts, floating matras, Nastaliq ligatures that change shape by position, and Bengali forms typeset in the 1990s. We fine-tune per script rather than trusting one multilingual model to cover all of them, and we publish the per-script numbers instead of one flattering average.
An audit trail that survives a regulator
Every field carries the page, the bounding box, the model version, the validation rules it passed, and who overrode it if anyone did. The trail is append-only. If a customs authority or an auditor asks why a value was accepted eighteen months ago, the answer is a query, not an investigation.
Where it runs
Managed by us, or inside your own VPC with your own keys. The self-hosted build is the same artefact we run — not a stripped variant — because a deployment story that only works in our cloud is not a deployment story.
What it does
- Schema-first field extraction
- Multi-script OCR incl. Devanagari & Nastaliq
- Rule + model validation with confidence gates
- Immutable per-field audit trail
The questions buyers actually ask
- Which scripts are supported today?
- Devanagari (Nepali, Hindi, Marathi), Bengali, Nastaliq (Urdu) and Latin. Others are a fine-tuning engagement, not a config flag — we will tell you the cost before you commit.
- What does 99.2% actually mean?
- Field-level accuracy at our default review threshold, across the largest production deployment, measured against human-adjudicated ground truth. Below the threshold the field is routed to review rather than guessed.
- Can we run it without egress?
- Yes. Air-gapped deployments are supported; model updates ship as signed artefacts you pull on your own schedule.
attatvasystems@gmail.com
Tell us what's breaking.
Send the problem, not a spec. We'll come back within one business day with an honest read on whether we're the right team — and who to talk to if we're not.
attatvasystems@gmail.com · Replies within 1 business day, Kathmandu time (UTC+05:45)