AIOrden: Secure Technical Notes Classifier

• 

  Local redaction: secrets become [[S1]] tokens in the browser and never reach the AI or the network.

• 

  GPT-5.6 classification: a chaotic note becomes a structured record; the credential shows only as [REDACTED_SECRET].

• 

  Records with an encrypted vault: credentials are stored with AES-256-GCM and masked as •••• until revealed.

• 

  Append-only audit trail: every secret reveal is logged as a secret.revealed event that can't be edited or deleted.

• 

  Learning library: resources and links classified with freshness tracking (current / review / outdated).

• 

  Project view: knowledge grouped by client. AIOrden documents itself as its very first project.

Inspiration

As a full-stack developer and IT freelancer, I accumulate technical knowledge everywhere: WhatsApp threads, loose notes, X and Grok, YouTube courses, bookmarks, RustDesk/AnyDesk sessions, client DVR configs, emails. When a client calls, I can rarely remember where I saved that remote-access setup or that credential. Searching wastes time — and pasting a secret into the wrong place is a real risk. I wanted one place that turns that chaos into structured, searchable knowledge without ever leaking a secret.

What it does

AIOrden is a privacy-first, installable PWA. You paste chaotic notes (or upload a .txt / WhatsApp export) and GPT-5.6 classifies them into clean, categorized records under a strict JSON contract — while every secret is stripped out in your browser before anything reaches the AI.

• AI classification of technical notes into a versioned JSON contract (infrastructure, credentials, commands, links, and more)
• Local secret redaction — credentials and keys become [[S#]] tokens client-side; the real values never touch the network
• Server-side leak scanner — if the model ever tries to return a secret, the entire output is rejected
• Human review with clarification questions before anything becomes a record
• Encrypted credential vault (AES-256-GCM) with an append-only audit trail
• Granular RBAC enforced in Postgres with row-level security
• Learning library & project registry — links and resources get freshness tracking; the app even documents itself as its own first project
• Installable PWA with an offline-ready app shell

How I built it — Codex + GPT-5.6

The part I'm proudest of: I didn't hand-write the code — Codex did. I acted as the orchestrator. For every task in my plan I wrote a bounded prompt and handed it to Codex CLI (running GPT-5.6, gpt-5.6-sol) as the sole implementer; I ran the quality gates and committed. AIOrden was built across ~59 gated Codex sessions — one per task — producing 318 passing tests (222 unit, 80 integration against the live cloud, 16 end-to-end).

The stack: a React + Vite + TypeScript frontend and a Supabase backend (Auth, Postgres with RLS, Deno Edge Functions). The classifier is an Edge Function that calls GPT-5.6 against a master prompt, and its output is validated by mirrored Ajv (client) and Zod (edge) schemas so malformed AI output can never reach the database.

My favorite detail: the same GPT-5.6 model that wrote the app also runs inside it as the live classifier.

Challenges I ran into

• Privacy vs. AI quality. The model needs context to classify well but must never see raw secrets. I solved it with client-side redaction ([[S#]] tokens kept only in React state) plus a server-side leak scanner — GPT-5.6 works on redacted text and is blocked from ever emitting a secret.
• Making an LLM reliable enough for a database. Free-form model output can't be trusted as a record. A strict, versioned JSON contract, mirror-validated on both client and server, turns GPT-5.6 into a dependable structured engine.
• Orchestrating an AI to build software. Driving Codex task-by-task — gated commits, fix rounds fed the exact failing test output, and a strict scope per session — was a new way of working, and it held quality across 300+ tests.

What I learned

That an agentic model can genuinely build a complete, secure, tested product when it's given tight scope, real gates, and honest feedback — and that "privacy-first AI" is an architecture decision, not a slogan: you design where the secret lives.

This isn't just a hackathon project — it's a tool I need every day.

Built with Codex + GPT-5.6 (gpt-5.6-sol): Codex wrote the code, GPT-5.6 classifies inside it.

Built With

• aes-256-gcm
• claude
• codex
• deno
• edge-functions
• gpt-5.6
• openai
• playwright
• postgresql
• pwa
• react
• supabase
• tailwindcss
• tanstack-query
• typescript
• vite
• vitest
• zod