Menu

Things I built

What I built. What reality taught me.

Each project began with something useful I wanted to make. The investigation emerged when the first version met reality.

Selected work

Start with something useful.

Each project began with a real product or customer problem. I learned the most by making the first version work, seeing where reality disagreed, and changing the system in response.

Working private system

Crispin

  • AI agents
  • Evaluation systems
  • Notion
Built
A Telegram assistant for organising work in Notion and Google Calendar, with focused agents and 300+ evaluations.
Reality taught me
A model can sound confident without completing the work. I learned to verify tool calls and state changes, then move guarantees into code.
How I built an agent I could trust
Working product

Eieye

  • Next.js
  • PostgreSQL
  • pgvector
Built
A full-stack AI intelligence product that collects HN stories, articles and comment trees, then turns them into a feed, weekly reports and cited research answers.
Reality taught me
Retrieval is only the start. Useful research depends on preserving source structure, inspecting both articles and discussions, bounding the agent, and checking every citation against evidence it actually read.
Making the HN hive mind queryable
v0.1 release candidate

mdReview

  • Go
  • Preact
  • TypeScript
Built
A browser interface for reading Markdown, commenting on documents or selected text, and saving feedback beside the original file.
Reality taught me
Comments can stay with a Markdown file without changing the document or moving it into another service.
Reviewing agent-made Markdown
Published experiment

Chat G&T

  • LoRA fine-tuning
  • Held-out evaluation
  • Structured outputs
Built
An interactive experiment comparing two ways of teaching a small AI model to turn questions into cocktail recipes.
Reality taught me
Fine-tuning learned the response contract more reliably than it learned judgment. It improved structure and removed most recurring prompt context, but did not establish that the answers were better.
Prompting vs fine-tuning in practice
Working pilot

Latently

  • TypeScript
  • Next.js
  • Hono
Built
A tool that gives AI assistants the right project context from Drive, Slack, Notion and files.
Reality taught me
Context is not simply a retrieval problem. Freshness, permissions, source truth and inspectability matter before embeddings or semantic search become useful.
Building reliable context for AI tools

Additional work

Founder products, automation and released applications.

Gather

Customer product · Provenance

A blockchain-backed product-provenance workflow for a furniture company.

→

Origin Thread

Founder product · Provenance

A clothing brand connecting physical garments to Cardano-backed provenance through NFC and QR technology.

→

MushroomRise

AI content automation

A Next.js workflow that turned scientific papers into short-form content and published through TikTok and Meta APIs.

Lock In Guru

Mobile product

A habit and discipline application released on the App Store.

Fliq

Mobile product

An educational discovery and microlearning application released on the App Store.

Mealll

Mobile product

A personalised recipe and meal-planning application released on the App Store.