Founding Engineer
Briskk · May 2025 – Present
- Cut typical latency for a production conversation-list API by approximately 85%—from 7.5-8.5s to 1-1.5s—using an Identify-to-Hydrate query architecture over approximately 71K conversations and 372K interaction logs; verified count, membership, order, and field parity across 26/26 benchmark cases.
- Built multi-tenant WhatsApp commerce systems across Python/FastAPI and Elixir/Phoenix, including structured LLM extraction, product discovery, pgvector retrieval, lead qualification, and human handoff.
- Built reusable deterministic and LLM-as-judge evaluation harnesses for intent, retrieval, hallucination, and response quality; replayed production-derived conversation scenarios and added grounded-response safeguards for price and product claims.
- Eliminated duplicate Al replies caused by BSP webhook retries through Redis-backed message-ID idempotency; reproduced a message processed four times and verified repeat deliveries were dropped before LLM execution.
- Reduced WhatsApp dispatch amplification by classifying permanent BSP failures and stopping futile retries; identified 80 daily rejections causing 240 wasted attempts, consolidated five drifted classifiers, and added a build-time guard for unclassified delivery states.
- Built key parts of a Meta click-to-WhatsApp sales journey spanning ad-context ingestion, product discovery, buying-intent qualification, store-visit booking, staff handoff, and visit-source attribution.
- Designed configurable lead-lifecycle and handoff workflows covering BSP adapters, staff/store assignment, ownership, retries, and audit events; consolidated three delivery implementations into one mutation-tested dispatch path.
- Operated production services on AWS using Docker and GitHub Actions, with Redis caching, rate limiting, and Prometheus/Grafana/Loki observability.
