Automatic Wiki
New knowledge is compiled into maintained, linked files that remain useful after the original conversation.
Selected work / Client knowledge and operations
AI-enabled operations system
Designed and delivered a messaging-first AI operating environment that turns instructions, voice notes and business context into coordinated work across connected tools.
The challenge
Knowledge, follow-up and recurring coordination were spread across conversations, documents and disconnected tools. The business needed practical AI support that could retrieve the right context, route work to the right specialist capability and keep operational actions under clear control.
My contribution
I led AI use-case discovery, solution scope and delivery priorities. My role was to translate business needs into structured requirements, shape the operating model and guide specialist technical work across agent roles, integrations, quality controls and adoption. This was solution and delivery leadership, not proprietary application coding.
What I built
The delivered system combines a source-aware RAG knowledge layer, persistent operating context, specialised agents and scheduled automations. An orchestrator routes work to focused agents for research, writing, planning, operational follow-up and QA. It creates one governed layer between a person's instruction and the connected systems that need to act on it.
Client-controlled Wiki and knowledge workspace
The system maintains an automatic Wiki as linked, human-readable Markdown files inside the client's own on-premise environment. The client owns these files and can browse, edit, reorganise or export them through a local knowledge workspace (e.g. Obsidian). They hold project context, operational notes, decisions and reusable knowledge. The retrieval layer reads the same local knowledge base when an agent needs context, rather than treating a public cloud store as the source of truth.
Client files
The knowledge workspace stays within the client environment as files the client can access and retain.
Knowledge workspace
People can browse, search and manage the Wiki through a local workspace (e.g. Obsidian), including a mind-map view where useful.
Local retrieval
Agents retrieve source-aware context from the same governed knowledge base when work requires it.
Operating system map
A clear operating sequence turns an instruction into governed business work. Each step has a distinct role and every connection follows the same path.
Message and voice
Instructions, notes and transcription
Knowledge workspace
Automatic Wiki, local files and mind-map view
Agent orchestration
Retrieves context and routes specialist work
Business actions
Calendar, inbox, projects and outputs
From message to action
Telegram is one interface, but the system can use the messaging platform a client already works in. A person can send a typed instruction or a voice note from their phone. Voice is transcribed into usable task input, while notes can be classified, tagged and recorded in the relevant knowledge or operations space. A plain-language instruction can then create a task, update a record or route work to the right specialist agent.
Capture
Turn messages and voice notes into structured notes, actions, projects and follow-up.
Coordinate
Route a request through the agent team and return the result to the same conversation.
Operational workflows
Calendar and tasks
Create or update calendar events, tasks, reminders, priorities and delegated work from a natural-language instruction.
Email operations
A scheduled inbox routine can identify messages that need action, prepare morning reply drafts for review and apply agreed classification or archive rules.
Project management
Where access is configured, create and maintain records in Notion and project-management tools such as Asana, Monday.com or Trello.
Business outputs
Create and update documents, Google Sheets, Google Slides and working files, then produce shareable exports such as PDFs.
Adoption and governance
The environment was designed for business use, with clear approval points, reusable skills, defined tool access and rules that keep durable operational knowledge separate from credentials. Connected systems act only within approved access boundaries. Email replies can remain draft-first for human review before sending. Its operating model is provider-agnostic. It can work across OpenAI, Anthropic, Google and open-weight models as well as different agent frameworks and integration tools.
Delivery at a glance
- Messaging-firstone conversation can trigger connected business work
- Multi-agentan orchestrator routes work to specialised delivery agents
- RAG-enabledsource-aware retrieval extends the business knowledge base
- Provider-agnosticworks across OpenAI, Anthropic, Google and open-weight models