Case Studies

How the work actually got done — the problem, the constraints, the architecture, the decisions and their trade-offs, and what came out of it. Written from the same Markdown the assistant answers from.

Ask AI about case studies

FeaturedDesigner and developer· September 2026 6 min read
Building Anjo AI: a portfolio that answers questions
A grounded, cited AI assistant with no database, running on Vercel for well under a dollar a month.
How this portfolio was designed as a retrieval-augmented AI assistant over Markdown — grounding rules, hybrid retrieval, a zero-infrastructure fallback, cost controls, and an evaluation suite — and what the trade-offs were.
- Markdown is the single source of truth: pages, search, chat cards, and the vector index all read the same validated files
- Hybrid retrieval (vector + BM25, reciprocal-rank fusion) over a committed JSON index — no database in production
- Every mode is grounded: retrieval finds nothing → the model is never called; answers cite numbered sources
- Next.js
- React
- TypeScript
- OpenAI API
- Zod
- Tailwind CSS
Read the case study
FeaturedCreator and maintainer· 2026 – present 5 min read
Making agentic delivery deterministic with Asterweave
An open-source framework that turns a free-form coding agent into a gated, resumable, evidence-backed delivery workflow.
Why autonomous coding agents become unreliable on real repositories, and how Asterweave answers that with a deterministic delivery graph, specialist agents with bounded write policies, an evidence contract, and hard safety rules around Git and GitHub.
- A deterministic graph (intake → analyze → challenge → plan → approve → implement → test → verify → review → submit PR) owns routing, retries, approvals, and evidence
- Thirteen specialist subagents with explicit tool sets and write policies; reviewers are read-only
- An evidence contract: gates pass only on real command output, diffs, or verified remote responses — never on model confidence
- Claude Code
- JavaScript
- Node.js
- Model Context Protocol
- GitHub Actions
Read the case study

Loading…

Building Anjo AI: a portfolio that answers questions

A grounded, cited AI assistant with no database, running on Vercel for well under a dollar a month.

How this portfolio was designed as a retrieval-augmented AI assistant over Markdown — grounding rules, hybrid retrieval, a zero-infrastructure fallback, cost controls, and an evaluation suite — and what the trade-offs were.

Markdown is the single source of truth: pages, search, chat cards, and the vector index all read the same validated files

Hybrid retrieval (vector + BM25, reciprocal-rank fusion) over a committed JSON index — no database in production

Every mode is grounded: retrieval finds nothing → the model is never called; answers cite numbered sources

Making agentic delivery deterministic with Asterweave

An open-source framework that turns a free-form coding agent into a gated, resumable, evidence-backed delivery workflow.

Why autonomous coding agents become unreliable on real repositories, and how Asterweave answers that with a deterministic delivery graph, specialist agents with bounded write policies, an evidence contract, and hard safety rules around Git and GitHub.

A deterministic graph (intake → analyze → challenge → plan → approve → implement → test → verify → review → submit PR) owns routing, retries, approvals, and evidence

Thirteen specialist subagents with explicit tool sets and write policies; reviewers are read-only

An evidence contract: gates pass only on real command output, diffs, or verified remote responses — never on model confidence