The work I'd want to be judged on.
Self-Auditing ML Decision Agent
Most ML models give you a prediction and stop there. This one doesn't. It pairs a gradient-boosted decision model with an agentic audit layer that flags its own uncertainty and verifies its reasoning against external tools — instead of trusting a single pass.
The audit layer is built on the Agent-as-a-Judge approach: rather than a plain LLM-as-a-judge that grades in one shot, the agent uses tool-use and real-world verification to catch the bias and shallow reasoning that single-pass judges miss. That self-checking behaviour is the part I'm most proud of, and it's what sets this apart from typical LLM-as-a-judge work.
Built end to end.
Real-Time Analytics & Insight Engine
A deployment-grade pipeline on a million+ retail transactions — cleaning, analysis, and an automated insight layer, served live with Streamlit and FastAPI.
SQL Analytics Dashboard
An industry-grade dashboard on a real relational dataset — advanced PostgreSQL at its core: window functions, CTEs, and the query patterns analysts actually use.
Comparative Backprop for NMT
An optimizer study (SGD, Adam, RMSProp, Adagrad) on an English–Hindi translation model built with Seq2Seq, LSTM, and attention. Best BLEU of 18.22 with RMSProp.
Quant Volatility & Risk Engine
A volatility and risk system for Nifty 50 using the GARCH family — regime detection, stress testing, and regime-aware allocation. Where I learned to respect real data.