Navneetha Rajan

Arizona's grid operators need to know which parts of the grid will fail before they fail, during Phoenix heat waves, evening EV charging surges, all of it. The problem is most forecasting models treat the grid as one big number. They miss that bus 47 affects bus 48.

So I built a Graph WaveNet that forecasts demand per bus, not per grid. 132 buses, 16 months of real demand and weather data, quantile predictions (p10/p50/p90) so operators know the best and worst case. The model learns which buses influence each other directly from data, no hand-tuned topology needed.

Then I plugged those forecasts into OpenDSS to run actual physics simulations: will voltage drop below safe limits? Will any lines overload? The dashboard shows operators exactly where to intervene and what to do.

Why Graph WaveNet over LSTM? The grid is a graph, not a sequence. Buses have spatial relationships that LSTMs can't see. The adaptive adjacency mechanism learns these relationships from data, which matters because the real wiring doesn't always match the official topology.

If you're a pharma market access analyst and you need to know "does Cigna cover Humira, and what hoops do I jump through to prescribe it?" today you're downloading PDFs from five different payer portals, opening each one, and manually comparing coverage rules across 50-page documents. For every drug. For every payer. Every time the policy updates.

I built a system that kills that workflow. Upload a payer's PDF, and a Gemini-powered extraction pipeline pulls out every drug, every prior auth requirement, every step therapy rule, and drops it into a structured database. Then you can search any drug and instantly see which payers cover it, what restrictions exist, and where the friction is. The standout feature was a heatmap showing prior authorization friction across payers and drugs. Judges stopped and stared at that during the demo because it made an invisible problem visible in one glance.

The pipeline is JSON to SQL to RAG. No vector store, no embeddings. Keyword-based retrieval into structured SQL queries, context assembled and injected into the LLM at query time. For a domain this structured, that approach outperforms semantic search because you need exact matches on drug names and HCPCS codes, not fuzzy similarity.