Incident Copilot

AI-Assisted Incident Triage System

TartanHacks 2026 Built at TartanHacks · Full-Stack · AI Systems

Designed and shipped in a weekend hackathon (36 hours).

Overview

Logs, alerts, tickets, and deploy notes in. A structured triage report out in seconds.

Hackathon Constraint

36 hours to ship a working AI decision-support system

Clear scope System-first Deterministic layers Strict outputs

TARTANHACKS 2026 · HACKATHON CLOCK

The Problem

Cognitive load, not missing data, drives resolution time.

Chaos Factors

Noisy logs Stacking alerts Customer reports

Manual Triage

Cross-reference signals Guess root causes Search past incidents

What I Built

Multi-signal ingestion
Deterministic severity scoring
Similar-incident retrieval
Schema-constrained triage generation

System Philosophy

Rules score severity. The LLM proposes hypotheses.

Speed Clarity Trust

MVP Scope

4 Signal Types Severity Scoring Hypothesis Generation Recommended Actions Confidence Estimation

Decision support, not automation.

SHIPPED AT TARTANHACKS 2026

Outcomes

36h from empty repo to working frontend and backend

TARTANHACKS 2026 · WITHIN TIME LIMIT

Schema-controlled output Deterministic scoring Incident memory

Skills: AI Systems Operational Intelligence Incident Response Full-Stack Development Hackathon Build

System Architecture

Built for operational speed

Frontend

Tailwind UI Triage layout Live AI binding

Backend

Express server OpenAI API Severity scoring Incident lookup

Strict Output Strategy

Strict JSON, no chat. Engineers read data, not prose.

AI Control Strategy

Engineering maturity beyond "call OpenAI and hope."

Schema Enforcement

Schema enforced in the system prompt, before parsing.

Resilient Parsing

Fallback structure. Hallucinated keys never crash the dashboard.

Trust Metrics

Confidence bars grounded by a deterministic severity score.

Tradeoffs & Decisions

Four calls under the 36-hour clock.

Retrieval

Vector database → keyword similarity, local dataset

Validate first, scale later

Interface

Chat interface → dashboard that answers on load

No time to chat in an outage

Interpretability

Opaque AI answer → numeric severity plus confidence bars

Calibrated trust

Actionability

Paragraphs of advice → step-by-step runbook actions

Mitigation starts immediately

Lessons from 36 Hours

Scope discipline matters more than feature volume.

Deterministic layers earn AI trust

STRICT PARSING · SCORING RULES

Structured output beats chat

DASHBOARDS · NOT CONVERSATIONS