Hippo

Context

Hippo started as an incident bot, but the goal was never "summarize chats."

The real job: convert live incident threads into structured decision objects the team can act on.

It generates 30-minute rolling summaries, extracts probable causes, and drafts RCA structure so the human team can move faster.

The deeper insight: incidents are recurring patterns — treat them as memory, not noise, and you can predict failure surfaces.

This project hardened my view that LLMs without structure, evals, and feedback loops become narrative generators — not operational systems.

Hippo is built to live inside workflows, not outside them.

Problem

• Incident response is chaotic: fragmented updates, repeated questions, and high context-switching cost.
• Decision latency increases when information is scattered across threads and calls.

What I Built

• Thread understanding → structured summaries (time-bucketed, decision-ready).
• RCA drafting scaffold (signals, hypotheses, timeline, contributing factors).
• Memory direction: learning from past incidents for pattern detection and prediction.

Guardrails & Constraints

• LLMs without structure become narrative generators — output contracts enforce decision-ready formatting.
• Summaries are time-bucketed to prevent context drift.
• Built for auditability: every summary links back to source threads.

Why It Matters

• MTTR improves when context is structured and continuously updated.
• Teams stop re-asking the same questions and move to decisions faster.

What's Next

• Pattern detection across historical incidents.
• Early anomaly surfacing before incidents escalate.
• Predictive failure surface modeling.