Ingested, normalized, and served equities/FX ticks with sub-100ms latency across services.

• 99.95% uptime across 90 days
• P50 < 45ms, P99 < 180ms end-to-end
• Storage cost reduced 22% via columnar partitioning

Backtesting and feature pipeline with reproducible runs and fast parameter sweeps.

• 10x faster sweep time via parallel runs
• Deterministic runs with artifact versioning
• Built-in slippage/fee models

Explored role-conditioned dialogue modeling for more accurate speaker intent classification.

• +4.2 F1 on benchmark conversations
• Inference cost reduced 18% via distillation
• Robust to speaker swaps

Unified metrics, logs, and traces with cost-aware sampling and clear SLO dashboards.

• 30% fewer on-call pages
• Clearer ownership and runbooks