Production AI for a mental-health platform

The problem

Sole ML engineer on a safety-critical product with zero ML infrastructure, where a wrong answer has real consequences.

The approach

• ▹Built 12 production systems from scratch: LLM serving with vLLM behind a stable API, a multi-tier real-time crisis-detection cascade, WHO/NICE-aligned clinical note generation, multilingual semantic search, and LLM intent routing, with adversarial safety evaluation and permanent regression harnesses.

The impact

• ✓>99% of traffic resolved in <100 ms
• ✓~99.7% crisis-detection accuracy
• ✓5.4× cheaper than external APIs

Trade-off, said plainly

The crisis-detection cascade is tuned for high recall over precision on purpose: for a safety-critical system, a false alarm is a much cheaper mistake than a missed one. That means more manual review than a precision-tuned model would need.

Stack