Distributed systems. Market data. Production AI.

A production AI system analysing call recordings at scale. A speech-to-text pipeline feeds a dynamic prompt engine where every analysis dimension — sentiment, objections, rep behaviour, drop-off signals — is user-configurable at runtime. Prompts are composed programmatically per configuration and dispatched to an LLM; results are post-processed into structured, actionable insight.

Recently shipped "Custom DIY Configurations": teams define their analysis parameters, because a collections team and a SaaS sales team need entirely different signals from the same calls.

Engineering core: sustaining high event throughput under variable STT latency while keeping LLM output quality stable across arbitrary user-defined configuration sets.

Event-driven telecom billing reconciliation at scale. Exactly-once processing across distributed Kafka consumers; designed for idempotency, fault tolerance, and late-arriving events.

Backpressure-aware consumption, deterministic replay orchestration for audit and recovery, and streaming Parquet export for downstream analytics.