Hacker
Karan Jakhar
@karanjakhar
Projects
- Top 50
Blastwave: on-call agents that work from the graph
Graph HacksOct 2026
When production breaks, the cause is rarely in the service that's alerting. It sits a few dependency hops away, in a deploy someone shipped 40 minutes ago, and the fix lives in the head of whoever handled the same failure months ago. On-call engineers rebuild that picture by hand from dashboards, git logs, Slack and ADRs. LLM agents don't do any better, because they see pieces, not connections, and forget everything between sessions. Blastwave is a team of three local-LLM agents (triage → investigator → comms) for SRE and on-call teams, and it uses FalkorDB as its context layer. Triage measures the blast radius with multi-hop traversals and PageRank. The investigator follows the dependency chain downstream to find recent risky deploys, matches past incidents, and pulls the fix procedure the team recorded last time. Comms pages the on-call engineer plus the person with real hands-on history, found from tickets, past resolutions and deploys, and posts a status update. Agents hand off through the graph: each one gets only an incident id and reads the shared state back from FalkorDB. When a human resolves the incident, the fix becomes procedural memory, and the next session, run by any agent, picks it up. Each decision is shown with the exact Cypher it ran and the subgraph it relied on, highlighted live in the UI. Everything runs locally on an 8 GB laptop GPU (Qwen3-4B), and the same tools are exposed over MCP.