A deterministic, zero-dependency prompt-pruning layer for long-running LLM conversations — expires stale tool state, collapses duplicate context, and proves it never drops a fact a later turn still depends on.

Production-style AI agent for automated contract-to-invoice pipeline with rule-based computation and validation

Automate the daily ingestion, transformation, and storage of stock market data using a secure, scalable pipeline.

Low-power fixed-point Softmax processing unit with LUT-based exponentiation and an 8-stage pipelined Mitchell logarithmic divider. Designed in Verilog and implemented on TSMC 180nm using Cadence tools. Achieves 1 output/cycle throughput and 8.9× speedup over sequential division. Patent under review.

A Python news monitoring pipeline that uses web scraping, LLMs and optional email alerts to keep up to date with specific topics.