Commind is a collaborative AI knowledge platform — an AI "team brain." It began from a multi-agent orchestration mandate and grew into a production system where agentic workflows plan, use tools, and act autonomously across a team's knowledge, meetings, and the web.
I came on as CTO and Lead Architect and designed the agentic layer: tool-using research agents, an autonomous meeting-to-action pipeline, and intent-routed retrieval — along with the engineering judgment to know when not to use an agent.
The agentic systems I architected
1 · Deep Research — a planner–executor loop
Decomposes a topic into sub-queries, runs multi-query web searches through a tool API, cross-references sources, and synthesizes a structured, cited report. Streams a transparent research trail — plan, search, cross-reference, synthesize — and degrades gracefully, returning partial findings on timeout rather than failing outright.
2 · Meeting Intelligence — perceive → reason → act
Autonomously ingests Teams transcripts via a delegated Graph API, summarizes them, extracts action items, resolves assignees against meeting-participant scope with close-match logic, and pushes tasks into Planner, To Do, Jira, and Linear. Cross-tool identity resolution maps participants to tracker user IDs.
3 · Intent-routed retrieval orchestration
The RAG reasoning layer: Query → Understand → Expand → Hybrid Retrieve → Rerank → Assemble → Generate. Classifies query intent — factual, exploratory, document-finding, expertise-location, temporal — to route strategy per query, blending dense (HNSW), sparse (BM25), and metadata filtering. Confidence thresholds and citation enforcement act as generation guardrails.
4 · Stateful conversational RAG agents
Document and presentation agents that maintain outline and source-ID state, recognize structural operations (add, reorder, modify), re-query the knowledge base on demand, and explain their own reasoning — answering "why this source?" with a relevance justification.
Signature architectural judgment
- Agentic when warranted, deterministic when better. I chose synchronous, debuggable pipelines and a one-report-per-conversation state machine over open-ended agent loops, and deliberately deferred a full agent/MCP layer until the RAG foundation was solid. The hardest call in agentic engineering is usually the one against using an agent.
- Guardrails by design. An ambiguous assignee results in no action taken, rather than a confident wrong one. Custom agents are scoped strictly to internal knowledge. Confidence gates and mandatory citations sit between retrieval and generation.
- Least-privilege by default. Delegated, user-controlled OAuth (PKCE) means agents act per user — never through admin service accounts. The system can only ever do what the person it's acting for could do.
The hardest call in agentic engineering is usually the one against using an agent.
Capabilities demonstrated
Have a similar challenge?
If you're building an agentic layer that has to plan, use tools, and act autonomously — safely, and with the judgment to know where an agent doesn't belong — let's talk.
Get in touch →