Computer vision & document ML
Layout analysis, detection, recognition architectures, and the evaluation harness that keeps them honest. If you have shipped something that had to work on photographs rather than scans, we want to talk.
Careers
Early, small, and unusually clear about the problem.
Most early companies are still looking for the question. We already have ours, and it is a hard one: teach a machine to read an Indian page properly. What we are short of is people, not direction.
Honestly
What you would be joining.
- ✓A defined technical problem with a measurable definition of better.
- ✓Work that will eventually touch people who never hear our name — the shopkeeper, the clerk, the student with a photocopied certificate.
- ✓Very few layers between what you decide and what ships.
- —Pre-revenue. We are building the model and the pipeline; there is no large customer base to inherit.
- —A small team. You would own areas outright, including the unglamorous parts of them.
- —No open listings today. We are not going to advertise roles we cannot yet fund properly.
Areas
Where we will need people.
Not job postings — an honest map of where the work is, so you can tell whether any of it is yours.
01 / VISION
02 / LANGUAGE
Indic NLP
Devanagari and the scripts after it. Language modelling for correction, tokenisation that survives conjuncts, and the judgement to know when a correction is a fix and when it is a fabrication.
03 / DATA
Data & evaluation
Building a training corpus of real Indian documents, synthetic augmentation for conditions we cannot collect enough of, annotation quality, and ground-truth sets we can defend.
04 / PRODUCT
Product & interface engineering
The capture and review surfaces. Making confidence legible, correction fast, and a technical system usable by someone who has never heard the word OCR.
None of these, but you're certain anyway?
Write anyway. The most useful people at this stage are usually the ones who do not fit a column neatly — and we would rather read a good letter than a matching CV.
How we work
A few things that are true here.
MEASURED
Claims come with evidence
Accuracy claims arrive with the document set they were measured on. This applies internally long before it applies to anything we publish.
UNGLAMOROUS
The boring stage matters most
Preprocessing outperforms model-swapping. People willing to do careful, unfashionable work on the parts that actually move the number do well here.
WRITTEN
Decisions get written down
Design choices are recorded with their reasoning, so a decision can be revisited later by someone who was not in the room.
Write to us.
Send something you have built and a few sentences on why this problem interests you. A link to real work is worth more than a formatted CV. Put Careers in the subject line.