Three active workstreams: price-anomaly detection over Northwind Commodities' trade blotter, document extraction for Orbit Supply covering roughly 30,000 pages a month of bills of lading and purchase orders, and retrieval over a contract library of 140,000 documents. Every one of them starts with an evaluation set, not a model. If you cannot measure it against a labelled sample the client agrees with, we do not ship it.
Full time — hybrid, three days a week on site
3-6 years shipping models other people depend on, rather than models that lived in a notebook
Bengaluru, India
Apply for this role
The work
What you will do
Build and maintain evaluation sets before any modelling work starts, including the adversarial and long-tail cases the client keeps quiet about.
Take the extraction pipeline from 91 to 97 per cent field-level accuracy across 14 field types, with a confidence threshold that routes the rest to human review.
Serve models inside a 400 ms p95 budget,
including retrieval, with a documented rollback to the previous version.
Instrument drift and volume monitoring, and define the trigger conditions for retraining rather than retraining on a calendar.
Explain to non-technical stakeholders exactly where the model will be wrong, before they find out in production.
You
What we look for
Python and PyTorch, plus practical work with transformer models — fine-tuning, quantisation, or serving them behind an API you kept up.
A real grasp of calibration, class imbalance and the difference between offline metrics and the decision the model actually drives.
At least one model deployed behind live traffic, with the monitoring and rollback path you built for it.
Healthy scepticism about published benchmark numbers, and a habit of reproducing them before quoting them.
The judgement to say a rules engine will beat a model on a particular problem, and to say it ea