Work

Problem

A high-traffic consumer platform was running a legacy ranking system on a slow batch pipeline. Personalization quality had plateaued and the system lacked meaningful handling for new users. The engineering team needed senior ML depth to redesign the approach.

Approach

Designed and built a two-stage retrieval and ranking architecture, replacing the batch pipeline with a near-real-time system. Integrated with the existing A/B testing infrastructure to enable a controlled rollout.

Problem

A content platform had built a processing pipeline using frontier model APIs as the backbone. As volume scaled, compute and API costs grew significantly and became a budget concern. Batch cadence created delays that blocked downstream product features.

Approach

Audited the full pipeline to identify where large models were being called unnecessarily. Designed a tiered processing architecture where lightweight models handle the majority of routine requests, escalating to larger models only when confidence falls below threshold. Migrated batch jobs to a streaming architecture.

Problem

A platform managing financial flows was experiencing systematic invalid activity that existing rule-based filters were missing. False negatives created financial exposure; overly aggressive rules were blocking legitimate users. Patterns changed frequently, making static rules inadequate.

Approach

Built an ML-based anomaly detection system combining unsupervised outlier detection with a supervised classifier trained on labeled cases. Designed the feature engineering layer to capture both per-transaction signals and aggregated account-level behavioral patterns. Implemented a feedback loop with regular model retraining.