Projects

Featured

• 

  Product · 2026

  Docket AI

  A WhatsApp-native assistant that lets small businesses create branded, numbered invoices, quotes, receipts, agreements and delivery notes by sending a message.

  The problem

  Small businesses sell on WhatsApp, but the paperwork — invoices, receipts, agreements — still happens somewhere else, often by hand.

  Why it's interesting

  The agent works inside a chat people already use: it asks for missing details instead of guessing, turns messy messages into structured documents, and keeps every document in a web workspace.

  Visit →

• Product · 2025

  AskHook

  A social-intelligence platform that analyses more than 500 short-form videos in minutes, distinguishes organic virality from paid promotion, and turns winning patterns into usable content decisions.

  The problem

  Trend data shows what moved, but rarely explains whether a pattern was earned, bought, or repeatable.

  Why it's interesting

  The agentic harness has to coordinate large-volume multimodal analysis, preserve evidence, and reduce noisy observations into decisions a strategist can defend.

  What I learned

  At this scale, orchestration and recovery paths matter as much as analysis quality. A useful agent needs to know what completed, what failed, and what evidence supports each conclusion.

• Product · 2025

  Axon · Neurochecklists

  An evidence-grounded neurology agent that searches and reranks more than 70,000 clinical checklist records to return structured, cited answers in under 2.5 seconds.

  The problem

  Clinicians and researchers need precise answers from a large specialist corpus without losing provenance, privacy, or time.

  Why it's interesting

  This is retrieval under real constraints: relevance is not enough unless the answer is fast, traceable to evidence, and shaped for a high-stakes workflow.

  What I learned

  Good RAG depends on the representation and reranking strategy around the vector store, not simply the choice of embedding model.

• 

  Product · 2024

  TrendMind

  An AI-assisted personal-branding platform exploring how intelligent systems can behave like a persistent content strategist.

  The problem

  Consistent, adaptive content across platforms is expensive to produce and hard to keep on-brand.

  Why it's interesting

  It's an applied experiment in agentic workflows, content intelligence and platform adaptation — one product touching every part of production AI I care about.

  What I learned

  The harness around a model — retrieval, memory, feedback, approvals — matters more than the model itself when reliability is the goal.

  Visit →

• 

  Open Source · 2023-04-01

  Code-critique

  An AI-powered CLI that reviews code inside your project before a pull request is opened.

  The problem

  PR review latency and context-loss make expensive humans re-explain the same issues on every branch.

  Why it's interesting

  The interesting problem isn't the review itself — it's compressing project context into something a model can act on without hallucinating.

  GitHub →

• Research · 2025

  Efficient Brain Tumour Segmentation (RAAGR²-Net)

  MSc research on reducing model complexity in brain tumour segmentation through parameter sharing and pruning — preserving segmentation quality while cutting computational cost.

  The problem

  State-of-the-art segmentation models are too heavy to deploy at the point of care.

  Why it's interesting

  A concrete case of the efficiency question I keep coming back to: how much can you take away from a model before its behaviour meaningfully degrades?