Knowledge Base
AI Integration into Processes
Challenge
Organisational knowledge is scattered across sources: transcripts, emails, chat exports, notes. Nobody can answer "what do we know about X?" without a manual search across multiple tools. The information exists — it just cannot be queried.
What we built
We built a two-layer knowledge store fed by the input pipeline: fleeting notes (append-only atoms, high recall, low cost) and evergreen wiki pages (synthesised from fleeting notes per entity, higher signal). Access is tiered: agents get a bearer-token API, humans get a web interface with Google OAuth + email allowlist. The web interface is self-hosted and sits behind a login gate.
Result
The inbox and notes layers are live. Agents query the knowledge base in real time during task execution. The wiki-page synthesis layer is in active development — as it ships, every entity across the portfolio will have a continuously updated reference page.
Tech used
- LLM note synthesis (compact AI model)
- Append-only fleeting notes store
- Evergreen wiki-page synthesis
- Google OAuth + email allowlist
- Bearer-token agent API
- Self-hosted behind login gate
delivered within 7 days