AI/ML Engineer at Citadel

Deployed XGBoost + LSTM hybrid models for intraday price move prediction, improving Sharpe ratio by 14-52% across 3 live trading strategies. Built real-time feature pipelines using Kafka + Bytewax processing 3.2M market events per second with p99 latency under 40 ms. Implemented MLflow + Kubeflow on AWS EKS, reducing model deployment time from 4 hours to 18 minutes via automated CI/CD. Developed anomaly detection system (Isolation Forest + Autoencoder) that flagged 97% of simulated spoofing patterns in live paper trading. Built agentic AI workflows using AWS Bedrock and Claude models with function calling, structured outputs, and autonomous decision-making for market intelligence and research automation. Optimized Redis cache management and Feast feature store retrieval strategies, reducing feature access latency. Created model monitoring dashboards with Evidently AI + Grafana, triggering automated rollbacks when data drift exceeded 0.15 PSI. Documented inference APIs via FastAPI + Ope

AI/ML Engineer at Zerodha

Designed batch inference pipelines computing daily VaR (95%/99%) and expected shortfall for 250K+ retail portfolios, eliminating manual risk calculations. Reduced feature engineering runtime by 44% using Polars + Spark structured streaming on 10+ years of NSE/BSE trade data. Fine-tuned a BERT-based model on 200K Indian financial news headlines, improving sector sentiment F1 from 0.68 to 0.79. Orchestrated Airflow DAGs with dbt + Snowflake, cutting ETL processing windows from 5.5 hours to 2 hours for daily market data. Built a RAG prototype using LangChain + FAISS to retrieve relevant disclosures from 10-K-equivalent reports for stock advisors. Automated data validation with Great Expectations + dbt tests, catching 41 data anomalies pre-ingestion in Q2 2024. Containerized ML training jobs with Docker + AWS ECR, reducing environment setup time from 3 days to 30 minutes for new team members. Presented model performance metrics weekly to product leads via Tableau dashboards, securing produ