Think of it as a smart filing cabinet for financial documents that maps relationships between companies and dates, then answers questions only with evidence it actually found.

• ▸Led enterprise KG-RAG product strategy, utilising Neo4j to extract complex financial entities from PDF data
• ▸Enforced AI reliability by designing 5-agent CrewAI pipelines with strict evidence-only JSON output protocols
• ▸Deployed Graphiti on Neo4j Aura, leveraging product telemetry and vector indexes for real-time visual graphs
• ▸Formalised financial domain modelling, building typed Neo4j edges to capture precise dates in enterprise filings

A Chrome extension that reads the financial article you are on, runs it through GPT-4, and gives you a plain-English summary with market charts in one click.

• ▸Synthesised real-time market analysis product pipelines, integrating GPT-4 APIs to deliver user news insights
• ▸Launched Chrome extension product features, triggering seamless LLM analysis across 200+ financial portals
• ▸Constructed interactive frontend UX with React and TS, embedding SVG charts for market data comparisons

You talk to your phone like a voice memo, and it turns that stream of thought into a structured journal entry with AI-generated reflections and insights.

• ▸Handled iOS scaling for 100+ concurrent OpenAI Realtime API connections with sub-second audio latency
• ▸Supervised backend product architecture via FastAPI and async PostgreSQL to power AI-driven journal synthesis
• ▸Prototyped iOS MVP roadmap and conducted local beta testing to validate core AI voice journaling UX workflows

An AI phone agent that schedules doctor appointments, designed to resist manipulation and reliably handle real conversations.

• ▸Built prompt-guarded voice flows for appointment booking, including zero-shot handling for names, dates of birth, and scheduling intents
• ▸Combined phonetic parsing with context-aware prompting to reduce hallucinated transfers and improve captured caller details

I trained an AI to play Pong from raw pixels until it could consistently learn winning behaviour through trial and error.

• ▸Implemented a convolutional DQN in PyTorch with epsilon-greedy exploration, target networks, and experience replay for stable Atari training
• ▸Optimised preprocessing and training with frame stacking, reward clipping, and CUDA-backed batches to improve sample efficiency

Employees could ask a question in plain English and get an answer from company documents instead of digging through PDFs themselves.

• ▸Built a retrieval-backed internal chatbot with LangChain, Flask, and GPT-4 to answer employee questions against company documents
• ▸Created ingestion and query pipelines that extracted PDF content, chunked knowledge, and served responses through a lightweight internal interface

Point a webcam at someone signing and the system translates the hand signs to text in real time.

• ▸Trained a transfer-learning pipeline on MobileNetV2 with MediaPipe landmarks to classify ASL gestures from live webcam input
• ▸Optimised inference with quantisation and efficient preprocessing to keep recognition responsive on edge hardware

I built a Connect 4 AI that looks several moves ahead and competed against more than 250 other agents in a class tournament.

• ▸Implemented Minimax with alpha-beta pruning and board evaluation heuristics to search several moves ahead under time limits
• ▸Tuned move ordering and scoring logic to cut decision latency and improve play against a large field of competing agents