An agentic finance platform that connects bank transactions, ledger entries, invoices, and emails to automate financial reconciliation and answer questions with cited evidence.
Integrated Elasticsearch 8 across seven indices to centralize financial data, with automatic fallback to local storage when the cluster becomes unavailable.
Built a GPT-4o-mini orchestrator with shared memory to route requests to specialist agents, including a reconciliation agent that matches bank transactions to ledger entries through four rule-based passes.
A reinforcement learning framework for training and comparing AI agents on Pokémon FireRed using game-state observations and gameplay interactions.
Developing a custom Gymnasium environment that extracts live game-state data from emulator RAM to provide structured observations for reinforcement learning agents.
Implementing and comparing PPO and DQN agents using PyTorch and Stable-Baselines3, with multi-seed experiments tracked through Weights & Biases and Matplotlib.