Acadmi is the academy for AI. It turns the usage data your AI tools already emit — starting with Claude Code and Codex — into levels, streaks and a live leaderboard for everyone on the team, and an adoption & ROI view for managers, so people get better at AI — without storing a single prompt, line of code, or tool argument.
Acadmi is a modular monolith: one backend, one agent. A reusable core handles ingest, storage, auth and the data API — and self-contained consumer modules turn that raw telemetry into something people use. New stakeholders plug in without touching the core.
Org-wide adoption & output across connected sources — trends, teams, locations, productivity, quality and ROI. The default landing for leads.
→ /analyticsThe fun layer: XP, levels, 15 trophy families, streaks, seasons and teams — everyday work scored automatically, no busywork.
→ the boardA config-driven signals registry parses OpenTelemetry into clean facts, stores them in Postgres, and serves a generic data API. Add a source, get a consumer.
→ ingest + APIThe /analytics site is a manager-gated dashboard that answers the questions leadership actually asks — who's adopting AI tooling, what it's producing, and whether quality holds — over rolling 7 / 30 / 90-day windows.
Illustrative figures — your numbers come from live telemetry.
Headline adoption & output KPIs across all connected sources.
How usage and output move over the selected window.
Adoption and output rolled up per team.
The same picture sliced by office / region.
Searchable per-person roster with drill-down.
Commits, PRs, reviews and active days.
Quality signals to read output honestly.
Where the activity is going — models, skills, repos.
Admin · manage the roster & ingest tokens.
Admin · org structure & project mapping.
Admin · value rollups for the program.
Admin · how far Acadmi itself has spread.
Managers only The analytics site sits behind a sign-in; the dashed views are revealed only for admin sessions.
Every session, commit, PR, review, skill and streak scores XP automatically — no check-ins, no manual logging. Climb levels, chase 15 trophy families, and reset the board each season for a fresh shot at the crown.
The agent runs on each person's own computer and reads the AI tools' own telemetry. Only counts and metadata are forwarded; anything sensitive is discarded at the source and again at the backend.
owner/repo), configured MCP server names/v1/traces — discarded entirelyA strict one-directional dependency rule keeps the platform core free of any product logic. The leaderboard and analytics are registered modules compiled into the same backend — not separate services.
A signals registry maps OpenTelemetry to stored facts. Extending to a non-Claude source is configuration, not a rewrite — and connectors/ pull from external APIs (e.g. Sonar).
Ingests Claude Code's own OTLP exports directly, so no upstream Anthropic API key is required. Complementary to the AI gateway, not dependent on it.
One ~18 MB distroless image. docker compose for local, Kubernetes for production — minimal or HA, with a DevOps sizing spec tuned for ~200 people.
Managers open the analytics site. Everyone else runs one install line, fires a prompt, and lands on the board. Raw data stays on the laptop — always.