Building reliable infrastructure for AI systems.

Problem

Products with legal weight, such as e-signature, need to verify who the customer is. The company relied on a paid external eKYC service.

What I built

• Datasets and detection models for ID cards and passports, plus MRZ reading.
• Checks for anything that makes a read unreliable: glare, blur, low light, occlusion.
• Checks for card validity: punched-hole detection for expired cards, cut corners.
• An internal Label Studio setup for labelling data.
• Production deployment with rate limiting, monitoring and maintenance.

Decision: no database

Each request is processed and answered immediately, and nothing is stored. When something fails I trace it by trace_id in Kibana. I was once asked to add a database to hardcode corrections for people whose names the OCR misread, for example when Vietnamese diacritics collide with printed patterns on the card. I declined. Only a few people a year were affected, and newer cards with a large QR code reduce misreads further. A ConfigMap file handles those cases, so the service stays stateless.

No database in the request path. Simplified overview, no internal names.

Problem

The company's GPUs are managed through Run:ai, but many features are missing from its UI and can only be reached through the API.

What I built

• A configuration management system that drives Run:ai through its API.
• LLM deployments with vLLM and SGLang, including custom images.
• Python libraries that integrate Logstash and APM to company standard and process in parallel, so streaming AI and LLM requests are not slowed down.