Solo-founded text-to-SQL SaaS built with FastAPI and React. Fine-tuned Gemma 4 using QLoRA (Unsloth) on curated text-to-SQL datasets, exported to GGUF and served via Ollama on a self-hosted NUC, achieving sub-second query generation with zero cloud dependency. Core differentiator is a semantic schema pruning layer that embeds schema metadata and retrieves only contextually relevant fields per query, cutting prompt token overhead by ~60% and significantly reducing hallucination rate. Unified analytics engine uses DuckDB to virtualize heterogeneous data sources (PostgreSQL, MySQL, CSVs) into a normalized SQL layer for non-technical users.

Built at ScarletHacks 2026, won Largest Market Opportunity. A drop-in FastAPI proxy that intercepts outbound LLM API calls, scanning and redacting Protected Health Information in real time using Microsoft Presidio and scispaCy before payloads leave the network perimeter. Full pipeline: request interception → PHI scan → selective redaction → forwarded completion → response reconstruction, requiring only an endpoint swap to deploy. PostgreSQL audit trail logs every detection event with entity type, confidence score, and timestamp for compliance reporting without storing raw patient data.

Distributed brain tumor detection system built with PyTorch and the Flower (flwr) federated learning framework. Trained on CT/MRI scans across 10+ simulated client nodes, raw patient data never leaves local nodes. Implemented custom classifier heads per client with a shared global aggregation strategy, improving global model accuracy by 12% over the federated baseline despite non-IID data distribution across nodes.

Interactive simulation tool that models future stock price distributions using Geometric Brownian Motion. Users input a ticker, time horizon, and number of simulation paths. Renders probability cones and percentile bands. Built with Next.js and deployed at index.kamakshya.xyz/monte-carlo.