WorkAI Knowledge & Operations System
A personal operating system, not a chatbot
AI Knowledge & Operations System
Converts fragmented AI conversations, project work and operational context into one searchable, actionable workspace — with agents that run the day on a schedule.
- Status
- In Use
- Category
- AI product
- Year
- 2026
- Role
- Conceived, architected and built as my own daily infrastructure.
The problem
Running several builds, a studio and a day job at once outgrew what any single AI chat could hold. Context evaporated between sessions, decisions lived in scroll-back, ideas died in notification trays — and every new conversation started with re-explaining who I am and what we'd already decided.
Constraints
- Capture had to work from a phone in seconds — friction kills personal systems faster than missing features
- Agents must run unattended across a 9.5-hour gap between host clock and local time
- Scrubbed summaries only — raw transcripts are never stored
- Own schema, own keys, no third-party note SaaS in the loop
The approach
Treat the AI's memory as infrastructure. Every working session gets distilled and ingested automatically — metadata and scrubbed summaries, never raw transcripts. Everything becomes searchable by meaning, not just keywords. On top of that memory sits an operations layer: a live control panel for capture, triage and planning; scheduled agents that run the daily rhythm; and a messaging assistant so the system is reachable from a pocket.
What was built
- Automatic session ingestion — hundreds of working sessions distilled, bucketed and stored with full-text and semantic search over all of them
- A key-gated live control panel (tasks, calendar, project status, inbox, command palette) built as an installable PWA with push notifications
- A messaging assistant with retrieval-grounded answers over the knowledge base, plus a scheduled morning standup
- A content pipeline: idea kanban with AI-drafted hooks and captions, staged from draft to ready
- A websites dashboard with health pings and a per-site admin-key vault (generate, rotate, hand to a client) — proven end to end
- A five-seat advisory-board workflow that convenes weekly on schedule to pressure-test decisions with deliberate candour
Contribution
Built with Supabase · Postgres · Edge Functions · Embeddings · Cloudflare Workers · Telegram Bot API · Web Push · Claude
Technical decisions
Store where agents are first-class
The prototype lived in a cloud-drive folder that agents could barely write to — which quietly reintroduced every manual step I was removing. Rebuilding on Postgres with a real schema made the memory something software can use, not something I maintain.
Summaries, not transcripts
The system ingests scrubbed session summaries and metadata only. Working memory shouldn't require archiving every raw word — a privacy decision that also keeps search results dense.
An OS, not an assistant
The unlock was scheduling. A shared database plus agents on cron turned good intentions into a system that runs whether or not I open it.
Evidence
Outcome
- In daily use as the grounding layer behind every other project on this site
- Hundreds of sessions searchable by meaning; nothing decided gets lost between conversations
- Scheduled agents fire reliably — including across the timezone gap, the hardest bug in the system
- Productised once as an experiment: a single sale at $199 over a local payment rail, deliberately not scaled yet
Current status
In UseIn Use daily — it is the operating layer this very rebuild was planned and verified through.
Lessons
A second brain is only as alive as the discipline around it — so the discipline became software. The rituals I used to forget are now cron entries.
Next stage
Selective productisation only if pulled by demand: one external sale exists, and multi-user hardening stays deliberately parked until there's a second.