Work
NL→SQL AI Agent
Machine Learning Engineer · Doowii
- –Engineered a production FastAPI system serving 50,000+ requests annually with 99.6% uptime; optimized end-to-end latency to 14.8s (p95).
- –Architected multi-model RAG pipeline in LangGraph with multi-node workflow orchestration, failover mechanisms, and strict latency guarantees.
- –Improved first-pass query accuracy from 62% to 78% on 600+ prod queries using hybrid retrieval (sparse + dense + re-rank; Pinecone) and schema-aware NER/entity linking.
- –Cut token costs by 31% and reduced p50 latency (10.3s → 8.0s) via Redis caching and an intent classifier routing simple queries to smaller models.
Multi-Tenant Dropout Risk
Machine Learning Engineer · Doowii
- –Scaled student churn prediction to 300+ universities via an automated model factory — validates data sufficiency, merges schemas, trains and deploys per-org XGBoost models via Vertex AI.
- –Designed a cold/hot feature store on BigQuery with point-in-time joins across heterogeneous tenant schemas; scored 3M+ records/day.
- –Enforced promotion gates (Precision > 0.6, F1 > 0.4); shipped isotonic-calibrated probabilities with SHAP-based feature importance for educator-facing interpretability.
- –Built multi-tenant monitoring (JSD drift + null-rate checks) with zero-touch onboarding for new tenants in <24h.