SentryAI won top-3 at a 48-hour hackathon: an AI assistant that watches a team's error logs and tells on-call engineers what broke, what changed, and what to do first. This is the honest breakdown — the architecture, the moments of panic, and what I'd change.
The Problem (20 Minutes to Frame It Right)
The hackathon theme was "developer productivity." Our first three ideas were rejected within an hour of discussion. The one that stuck came from a complaint I'd heard from a real teammate:
"I get paged at 2 AM. By the time I've figured out which deploy, which service, and which error started the spike, 40 minutes are gone."
The insight: on-call engineers don't need more dashboards. They need a diagnosis.
The Architecture (What Actually Shipped)
┌─────────────┐
error log stream ─▶ Ingest API │ (Node, streaming)
└──────┬──────┘
▼
┌────────────────┐
│ Time-window │ group errors by
│ grouping │ (service, error class, 5min)
└───────┬────────┘
▼
┌─────────────────┐
│ Anomaly check │ spike in rate vs 7-day baseline
└───────┬─────────┘
▼
┌───────────────────────────┐
│ LLM diagnosis │ prompt: error context +
│ (system: on-call expert) │ recent deploys + service graph
└───────────┬───────────────┘
▼
┌────────────────┴────────────────
▼ ▼
Slack alert with "Root cause hypothesis +
severity & timeline first action" card
Deliberate simplicity. No custom model, no vector DB in the end (the first night we added one; the second night we deleted it — the context was small enough to stuff directly into the prompt).
The Three Decisions That Decided It
1. We built for the judges, not the docs
The demo needed to work in 90 seconds. So we pre-seeded a realistic incident: deploy at 14:02, error spike at 14:05, one service. The judge watched a page arrive, saw the alert, and read a diagnosis that named the deploy. We rehearsed that exact flow 11 times. Every hackathon demo that fails fails in rehearsal, not on stage.
2. We cut features with a timer
Hour 20 feature list: dashboard, alerts, diagnosis, team assignment, cost tracking, mobile. Hour 24 reality check:
Keep: alerts + diagnosis (the core promise)
Cut: dashboard (a Slack alert IS the interface),
cost tracking, mobile
Shrink: team assignment → a /mention in the alert
[!TIP]
- The cut list is a hackathon skill. The rule we used: anything that doesn't make the demo sentence true gets cut. Our demo sentence: "When your service breaks, SentryAI tells you what happened and what to do, in under a minute."
3. The LLM prompt was the product
The diagnosis prompt took 4 iterations. The winning version:
You are a senior on-call engineer.
Given: (1) error burst summary, (2) recent deploys, (3) service dependencies.
Respond in EXACTLY this format:
SEVERITY: <P1|P2|P3>
ROOT CAUSE: <one sentence, name the change if evidence supports it>
EVIDENCE: <2 bullets, specific>
FIRST ACTION: <one concrete command or step>
Constraints: If evidence is insufficient, say so — never invent a cause.
The format constraint and the "never invent a cause" line were the difference between a credible tool and a confident hallucination machine. Judges noticed.
The 3 AM Moment
Hour 33: the ingest pipeline silently dropped ~30% of log lines under load. We had 15 hours left. The naive fix (better pipeline) was a day of work. The honest fix: the demo doesn't need throughput, it needs truth. We rate-limited the demo data, added a note to the README ("ingest is not load-tested; diagnosis path is"), and moved on. Shipping a smaller true thing beat a bigger broken one.
Results
- Top 3 of 24 teams (judge scoring: problem, execution, demo, pitch)
- A judge message afterward: "This is the first entry that felt like a product, not a demo."
- Our repo became the reference I point to when people ask what I've built
What I'd Change
- Start with the demo sentence, not the idea. "When X happens, our tool does Y in Z time." Write it before hour 3.
- The README is marketing. Ours was written at hour 40 and honestly carried us — problem, architecture, demo video, honest limitations.
- Sleep is a feature. We rotated a 90-minute sleep shift. The team that was up 40 hours straight pitched like it.
- Choose boring infrastructure. A free LLM API tier, a managed DB, no custom infra. Hackathons are for the idea, not the ops.
Discussion (1)
Sign in or create an account to join the discussion.
2mo ago
The "build for the judges, not the docs" lesson is the real hackathon wisdom. We skipped this last year and it cost us.