Ahmed Hmeid Contact
All work 01 · AI delivery platform

Bynary

AI made building easy. Bynary makes building right.

Open bynary.hmeid.uk

The live demo is briefly offline while it moves to new hosting.

What it is

Bynary is an end to end product lifecycle harness. A feature starts as a card on a roadmap, goes through an AI planning conversation where a human makes the calls, then Claude Code builds it inside a cloud sandbox. Bynary tests it, reviews it, deploys it to a preview URL, records a walkthrough video, writes the docs and opens the pull request.

  • 5phases per cycle: plan, build, review, document, ship
  • 714commits over five months
  • 64tables, with 186 row-level security policies
  • 3AI review agents: security, accessibility, design

Key features

  • Ideation and roadmap

    A product-manager persona helps shape features, but a feature can only reach "Ready" with a persona and a user-value statement. Now, Next, Later kanban that refuses to promote a feature ahead of its dependencies.

  • Planning with consent

    An engineering-lead persona plans each feature in chat, with interactive pickers for stack, fonts, colours and infrastructure. Tech decisions need explicit sign-off before anything is written.

  • UI generation

    HTML wireframes per feature, stitched into one click-through preview, plus a generated design system and component library that later builds must reuse.

  • Autonomous build

    Each feature gets a fresh Claude Code session in an E2B sandbox, with the whole roadmap in its context. A type-check gate runs between features and a git push checkpoints each one.

  • Auto test and preview

    Runs the project's own build, lint and tests, checks every acceptance criterion with AI, deploys a preview on Coolify and fixes failed deploys itself, up to three rounds.

  • AI review, then human review

    Security (OWASP), accessibility (WCAG 2.2 AA) and design review agents raise issues before a person reviews each task with diffs, logs and recorded Playwright walkthroughs.

  • Docs and merge

    Generates user docs, release notes and technical docs, then opens a GitHub pull request with the review outcome, file stats and a manual-steps checklist.

  • Traceability and cost

    Audit log, decision log, a knowledge map from task to file and function, and the cost of every AI call, per cycle and per stage.

In action

Illustration of the roadmap board: Ready features move from Now into a locked In Cycle lane when a cycle starts
Illustration of the planning chat: the AI proposes tasks and done-when criteria, and nothing is built until a human approves
Illustration of a live build: Claude Code works in an E2B sandbox through build, validation and AI review phases
Illustration of the review screen: per-task acceptance, security/accessibility/design review agents and a recorded preview walkthrough
Under the hood

The parts that aren't obvious

  1. The database is the queue

    API routes only set pending flags. A separate worker claims jobs with an atomic UPDATE so only one worker wins, with heartbeats, stale-job recovery and graceful shutdown. The same loop runs builds, docs, previews and GitHub sync.

  2. A scope guard inside the sandbox

    A Claude Code PreToolUse hook blocks edits to files outside the task, and a git pre-commit hook catches anything that slips through Bash.

  3. Sandboxes that cannot leak money

    Each sandbox is registered in the database before it is created, so a crash still leaves a record for the reaper. Pooled sandboxes rotate before the one-hour cap, and long builds resume from git checkpoints.

  4. Regressions, not noise

    A toolchain baseline separates errors that were already there from ones the agent introduced, so fix rounds are not wasted on old problems.

  5. A supervisor for the agents

    An always-on monitor spots stuck or looping builds, injects hints, and pauses for a human when it hits something it cannot fix, like a missing secret.

Built with

App
  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind CSS v4
  • shadcn/ui
  • Plate editor
  • Mermaid
AI
  • Claude Sonnet 4.6
  • Claude Opus 4.6
  • Claude Haiku 4.5
  • Claude Code
  • MCP
Infrastructure
  • E2B sandboxes
  • Supabase
  • Coolify
  • GitHub App
  • Docker
Testing
  • Vitest
  • Playwright
  • Playwright MCP