Where the work has actually been done.
Four long-running programs - health, conservation, identity and industry - that taught us the rest of the stack. Each one has shipped, has results, and is still running.
Catching oral cancer from a phone camera.
India sees roughly one-third of the world's oral cancer cases - and most of them are caught too late, in villages where the nearest oncologist is fifty kilometres away. With the Asian Head & Neck Cancer Foundation, LENS trained a model on tens of thousands of intra-oral images to flag pre-malignant lesions on a low-end Android phone, in dental camps run by ASHA workers.
- 01Sensitivity 96.2%. Validated against biopsy on a 2,400-patient cohort.
- 02140+ camp sites. Across Maharashtra, Tamil Nadu, Karnataka and Uttar Pradesh.
- 03Runs offline. Inference on-device. Sync when network returns.
- 04Explainable. Lesion heatmap returned with every flag - the screener sees what the model saw.
Recognising mountain gorillas by face.
The world has fewer than 1,100 mountain gorillas left. Tracking each individual matters - for poaching response, for veterinary intervention, for census. LENS retrained the same face-recognition substrate that runs at airports on a population of 1,063, working with field biologists and the Global Initiative Against Transnational Organized Crime.
- 01Re-id 94% on the closed population, including juveniles.
- 02Camera-trap pipeline. Ingests hundreds of thousands of frames a week.
- 03Anti-poaching alerts. Cross-referenced with patrol logs in real time.
- 04Open dataset. Anonymised features released to the conservation research community.
Fingerprinting infants at birth.
An infant fingerprint is roughly a tenth the size of an adult's, on skin that hasn't yet hardened. Conventional sensors don't see it. With KIET and ZKTeco, LENS built an optical pipeline - custom optics, custom contrast model, custom matcher - that captures and matches infant prints from day zero. A verifiable identity from the first hour of life.
- 0114,200+ infant cohort across hospitals in Delhi-NCR and Bengaluru.
- 02Match-rate 97.8% at twelve months against the day-zero print.
- 03Vaccination linkage. No more lost immunisation records when families move districts.
- 04Anti-trafficking. An identity that cannot be stolen.
MIDAS · reading a furnace floor.
JSW Steel asked a hard question: can a camera tell us a conveyor tyre is going to fail before it does? MIDAS - the LENS industrial model - reads billet, slag, conveyor and mould across visible and thermal sensors, flagging anomalies before a metallurgist would. In production at Vijayanagar across 22 lines.
- 01Predictive failure. Tyre wear, mould blockage, slag overflow flagged 4-36 hours ahead.
- 02False-positive <0.4%. Operators trust the alert.
- 03Multi-sensor fusion. Visible and thermal reasoned together, not stitched at the dashboard.
- 04Explainability surface. Every alert opens to the contributing frames and a trend the metallurgist can confirm.
Run a program with LENS.
If your problem looks like one of these - long-horizon, sensor-rich, high-stakes - we'd like to hear about it.