Dev Builds AI On-Call Assistant That Remembers Past Incidents — 93% Accuracy With Just 10 Histories
Front_Engineering591 · reddit · 2026-11-08
A developer built Incident Responder using Python, FastAPI, pandas, Groq, and Vectorize's Hindsight as a long-term memory layer. It ranks fixes by historical success rate, recalls related incidents for context, and lets an LLM explain (not compute) recommendations. Key decisions: all statistics computed outside the LLM, plus a shrinkage step so 1/1 success doesn't outrank 8/11. On synthetic data, top-suggestion accuracy went from 20% with zero history to 93% with 10 incidents and 100% at 50. The author candidly lists limitations (synthetic data, type-only alert matching) and asks SREs for feedback on trust, memory content, and adoption.
More from coding & agent
- Dimillian points to OpenAI Showcase for prompts, plans original demo-and-prompt series — Dimillian · 2026-11-08
- Testing ICE Constrained Decoding: Enforce Constraints at Generation, Not After — manvigupta03 · 2026-11-08
- ResearchBoost Agent speeds up real research workloads 2.51x on average across 12 tasks — YouJiacheng · 2026-11-08
- Built an interactive 3D optics course for GCSE physics entirely with Claude Opus 5.5 — techartist_ · 2026-11-08
- Perplexity CEO uses an agent to unmask anonymous recruiter candidates and reach them directly — AravSrinivas · 2026-11-08
- Alibaba's Qoder Launches Stealth Model Qoduck, Free for Individual Users — Aiden_Tech_Ai · 2026-11-08