AI Peer Group: Turning Group Conversations into Searchable Organizational Intelligence

17 April 2026#worksona#portfolio#mcp#knowledge-graph#peer-learning#multi-tenancy

Peer learning groups generate real value during live sessions. That value typically vanishes afterward — transcripts land in Drive folders, insights get lost in Slack, and three months later nobody can recall who recommended what or why. AI Peer Group is built on the premise that this doesn't have to be true.

What It Is

AI Peer Group is a platform for creating and managing AI peer learning cohorts. Group leaders configure groups with custom landing pages, manage member applications, set up meeting schedules, and connect recording infrastructure. Members join via public landing pages, attend sessions, and access accumulated group knowledge through the platform or directly through their IDEs and AI assistants.

The core system processes every session transcript through an 8-step knowledge graph pipeline, then exposes the resulting structured knowledge through a Model Context Protocol server. A developer in Cursor can ask "what did the group discuss about RAG architectures?" and get cited, speaker-attributed answers from actual session transcripts — without leaving their editor.

The platform also handles the full operational surface of running a peer group: email notifications across the member lifecycle, Google Drive import for manual transcript uploads, Recall.ai bots for automated session recording, a Stripe-backed Free/Pro/Enterprise tier model, and a role-aware admin dashboard for group leaders and platform operators.

Why It Matters

Most platforms that serve communities stop at "store the content." AI Peer Group treats every conversation as raw material for structured, queryable knowledge. A group that meets 50 times generates thousands of mentions of tools, frameworks, concepts, and decisions. Flat text search returns individual mentions. The knowledge graph returns: "this framework was discussed in 12 sessions across 8 months, it's consistently related to these other concepts, and here are the specific speakers and passages."

The zero-effort automation is what makes this viable in practice. The full pipeline from live meeting to searchable knowledge requires no manual steps from group leaders. A Recall.ai bot is scheduled automatically, joins the meeting, produces a speaker-attributed transcript, and triggers the knowledge graph build via webhook — all without anyone pressing record or uploading a file.

The MCP distribution layer means the knowledge doesn't sit behind a login screen members have to visit. It surfaces wherever they already work.

How It Works

flowchart TD
    A[Recall.ai Bot\nauto-scheduled] -->|done webhook| B[Transcript Converter\nspeaker attribution]
    C[Google Drive\nweekly cron scan] -->|txt/docx/vtt/pdf| B
    D[Manual Upload] --> B

    B --> E[KG Pipeline\n8-step orchestrator]
    E --> E1[Chunk ~400 tokens\nspeaker-aware]
    E1 --> E2[Claude: entities + relations\nperson/concept/tool/framework]
    E2 --> E3[OpenAI: embeddings\n1536d vectors]
    E3 --> E4[Upsert nodes + edges\nincrement mention_count]
    E4 --> E5[Create citations\nchunk to node provenance]

    E5 --> F[(Supabase PostgreSQL\n+ pgvector\nRLS group isolation)]

    F --> G[MCP Server\n9 tools + 3 resources]
    F --> H[REST API v1\nBearer token + scopes]
    F --> I[Admin Dashboard\nKG explorer + analytics]

    G --> J[Claude Desktop / Code\nCursor / Windsurf]

The knowledge graph pipeline runs 8 steps per session: fetch transcript, chunk into ~400-token speaker-aware segments, embed chunks via OpenAI, extract entities and relations via Claude, embed nodes, upsert nodes and edges with merge-on-conflict logic, create citations linking chunks to the entities they evidence, and mark the session as indexed.

Search combines vector similarity over both chunks and nodes with 1-hop graph expansion — finding not only text that matches the query semantically, but also structurally related entities that the group has connected to matching concepts across sessions.

Multi-tenancy is enforced at the database level via PostgreSQL Row-Level Security. Every knowledge graph table carries a group_id and has RLS policies that make cross-group data access structurally impossible, not just filtered in application code.

The MCP server exposes 9 tools including search_knowledge_graph, get_entity, explore_graph, and import_session, plus 3 resources providing ambient group context.

Where It Fits in Worksona

AI Peer Group applies the same knowledge infrastructure philosophy as Meshwork to a specific social context: the peer learning cohort. Where Meshwork is a general-purpose organizational knowledge backend, AI Peer Group is a full product — with group lifecycle management, email, billing, and a public presence — built on top of the same core knowledge graph and MCP distribution pattern.

The platform's strategic bet is that the winning distribution channel for group knowledge is not a dashboard members visit periodically. It's their existing AI-powered tools, where the knowledge surfaces in the flow of work. MCP is the protocol that makes this possible today.

Live: ai-peers.com

𝕏 Post