A small studio building document intelligence, computer vision and agents, taken from prototype to production with the evaluation to back it.
OCR, layout-aware extraction and long-context understanding, grounded in the source.
Detection, tracking and recognition tuned to run on embedded and edge hardware.
Tool-using LLM systems with bounded retries and replayable evaluation.
An in-house program for the failure cases standard models miss.
An AI assistant that triages cases, surfaces evidence and drafts decisions for an officer to sign off, on top of document intelligence for KYB and screening NLP across adverse-media, sanctions and PEP risk.
Detection, tracking and plate recognition across several camera feeds on one constrained edge device, with the recogniser fine-tuned for Arabic-script plates.
Mapping each field on an unfamiliar government form to the label that describes it, so a filled form reads back as structured data. Backed by a EuroHPC grant.
