How it works
The bot is a guided intake and routing system — not an autonomous operator. It never changes infrastructure, purchases, or production systems.
The flow
- Discover — someone mentions @aicoeBot in the AI CoE space, uses a slash command, or the proactive scanner spots a likely cost/quality problem (below).
- Intake — the bot opens a private DM and asks a few targeted questions at a time (buttons/dropdowns where answers are finite). Conversation state is durable and resumable.
- Structure — the intake is validated into a schema-checked record; deterministic code — never the LLM alone — drives what happens next.
- Route & ticket — a deterministic router picks the best-matching available advisor from the roster; an internal ticket (AICOE-XXXXXX) is created, tied to the requester's verified Google Workspace identity.
- Hand off — the requester gets a confirmation; the advisor gets a DM with the structured summary and evidence package.
Proactive scanning (cascade)
- A very low-cost classifier (muse-glimmer-30b, Fireworks) reviews new monitored-space activity about once a minute when there is new data.
- Flagged candidates go to a stronger triage model; only high-confidence agreement produces a suggestion.
- Guardrails: in-thread card only (Start private intake / Not now / Dismiss), at most one suggestion per person per day, thread- and user-level suppression. Calibrated on real messages — see the analysis.
Spaces we monitor
| Space | Status | Notes |
|---|---|---|
| AI CoE — Go-To for AI Questions ↗ | Monitored · bot home | The dedicated AI CoE space: the bot lives here, takes mentions and slash commands, and posts proactive suggestion cards. |
| The AI Watercooler ↗ | Monitored | Free-form AI discussion. Scanned by the cascade; suggestion cards allowed with the standard guardrails. |
| General ↗ | Monitored | Company-wide space. Scanned by the cascade; suggestion cards allowed with the standard guardrails. |
| AI Spend ↗ | Calibration source | Cost-focused space; used for calibration data (quiet — 2 messages in the sample window). |
Proactive suggestion cards may appear in all monitored spaces with identical guardrails
(per-person-per-day cap, suppression, dismiss). AI Spend is a calibration source only.
Where things live
- State (cases, tickets, audit): Postgres on the enterprise Supabase org — the system of truth.
- Recall (conversation memory, entity knowledge graph, reasoning traces): self-hosted Neo4j agent memory.
- Runtime: Hermes Agent on AWS Lambda microVMs, models via AWS Bedrock; classifiers via Fireworks AI.