🚨 Mayday — AI Incident Commander

A swarm of 6 Gemma 4 31B agents triages your production incident in seconds — running on Cerebras.

Built for the Cerebras × Google DeepMind Gemma 4 24-Hour Hackathon. Mayday turns an alert — a dashboard screenshot, raw logs, and a runbook — into a root-cause diagnosis, a safe remediation command, and a ready-to-paste Slack update. The whole investigation finishes in ~2.5 seconds because every agent runs on Cerebras at 1,000–2,400 tokens/sec.

Multimodal · multi-agent · speed-native. Targets Track 1 (multi-agent + multimodal), Track 3 (enterprise / incident response), and Track 2 (People's Choice).

---

The swarm

Stage	Agent	Codename	Does
1 ‖ parallel	👁️ Vision Analyst	OPTIC	Reads the alert dashboard screenshot (multimodal)
1 ‖ parallel	📜 Log Analyst	TRACE	Finds the smoking gun in the logs
1 ‖ parallel	📚 Runbook Retriever	ARCHIVE	Matches the relevant runbook section (and rejects distractors)
2	🧠 Root-Cause Analyst	SHERLOCK	Synthesizes a leading hypothesis with a causal chain
3	🛡️ Skeptic	DEVIL	Adversarially challenges the hypothesis before it's trusted
4	📣 Incident Commander	MAYDAY	Issues the final structured decision: severity, timeline, safe command, Slack update

Stage 1 fans out in parallel — the moment Cerebras speed compounds. Then SHERLOCK → DEVIL → MAYDAY run in sequence. Everything streams live to the UI.

Why Cerebras + Gemma 4

• Multimodal: OPTIC reads a real Grafana-style dashboard via Gemma 4's image_url (base64) input.
• Structured outputs: MAYDAY emits a strict-JSON decision (response_format: json_schema, strict).
• time_info: every response carries real timing — we surface live tokens/sec + TTFT.
• Speed race: a built-in side-by-side running the same model — Gemma 4 31B on Cerebras vs the same model on a reputable full-precision (bf16) GPU provider (W&B Inference via OpenRouter, pinned for a fair + reproducible comparison; not an AI-chip competitor), measured identically on both lanes — typically ~20–40× faster on output throughput.

Proof — it generalizes

Mayday is validated on 10 real-world incidents across 3 domains, each with its own dashboard, logs, runbook, and hand-written ground truth:

• Ops / SRE (6): DB connection-pool exhaustion, Redis cache stampede, a memory-leak OOM storm, a bad feature-flag rollout, a downstream / 3rd-party outage, and a metastable retry storm.
• Security / SOC (2): distributed credential stuffing and data exfiltration via a compromised credential.
• FinOps / Cost (2): an over-provisioned ASG after a bad Terraform apply, and an orphaned idle GPU cluster.

The same 6-agent swarm handles all of them — and picks the right class of fix each time across six remediation classes (rollback · disable-the-flag · fail-over · shed-load · contain · cost-fix), which is the test of real reasoning rather than pattern-matching. scripts/eval.mjs runs every incident end-to-end through the live swarm and grades the diagnosis with a Gemma-4 judge:

10 / 10 correctly diagnosed · avg ~2.3s per incident · peak ~3,000 tok/s. → docs/EVAL.md

Run it

npm install
cp .env.example .env.local      # then paste your CEREBRAS_API_KEY (csk-...)
npm run dev

Open the app, click load sample, then 🚨 DISPATCH SWARM. Or drop your own Grafana/Datadog screenshot + logs + runbook.

Environment

Var	Required	Notes
CEREBRAS_API_KEY	✅	From cloud.cerebras.ai. Model is gemma-4-31b.
BASELINE_API_KEY / BASELINE_BASE_URL / BASELINE_MODEL / BASELINE_LABEL	optional	A real GPU provider for the speed race. Without it, a clearly-labeled representative baseline is shown.

Scripts

node --env-file=.env.local scripts/test-cerebras.mjs   # smoke-test the API (chat, streaming, image, structured)
node --env-file=.env.local scripts/eval.mjs <baseUrl>  # run the 7-incident accuracy eval -> docs/EVAL.md
node scripts/gen-dashboards.mjs                         # regenerate all scenario dashboard PNGs
npm run build                                           # production build

Architecture

app/
  page.tsx                  server entry → <Console/>
  api/incident/route.ts     SSE: runs the swarm, streams every token
  api/race/route.ts         SSE: Cerebras vs GPU speed race
  api/{config,scenario}     UI bootstrap
lib/
  cerebras.ts               OpenAI client → Cerebras; streamAgent + runCommander (structured)
  orchestrator.ts           the 4-stage fan-out, emits StreamEvents
  agents.ts / roster.ts     system prompts (server) + roster metadata (client)
  race.ts                   speed-race runner
  data/                     agent-prompts.json · scenario.json
components/                 Console, AgentCard, SpeedHud, CommanderPanel, SpeedRace
scripts/gen-dashboard.mjs   builds the Grafana-style sample dashboard
docs/                       DEMO_SCRIPT.md · SOCIAL.md · SUBMISSION.md

Demo & submission

See docs/DEMO_SCRIPT.md (60-second video), docs/SOCIAL.md (X launch thread), and docs/SUBMISSION.md (full writeup).

---

Built on Gemma 4 31B running on Cerebras. #Gemma4