Hacker

Pradeep Konda

@pradeeptrain

Joined Oct 2026AI Engineer at TechnumenIndia

Projects

  • Top 50

    Ariadne — an on-call agent whose context and memory are a FalkorDB graph

    Graph HacksOct 2026

    When an alert fires, the on-call engineer has to answer three questions the company already knows the answers to: what changed, what else breaks, and who owns it. The deploy history, the dependency graph and past incidents all hold the answer, but as rows nobody can follow under pressure. Google's SRE book attributes roughly 70% of outages to changes in a live system, and Cloudflare's 18 Nov 2025 outage is a recent example: the trigger was a permissions change deployed at 11:05, and the investigation spent two hours on a downstream service and a suspected DDoS first. Ariadne is for on-call and SRE teams. An alert arrives (including through a real Prometheus Alertmanager webhook), and Ariadne returns the culprit commit, every affected service, the owner to page and the runbook, as a path through a FalkorDB graph of services, deploys, commits and people. Ranking is computed in Cypher, never asked of the model: Claude only chooses which graph tool to call and writes the explanation. Then it remembers. Each resolved incident is written back as a plan keyed on the structural shape of the situation (metric kind, service tier, fan-in, fan-out, whether something deployed upstream, severity), so the next alert with the same shape, even on a different service, recalls it. Measured over 20 alerts with memory withheld vs available: 1.6× cheaper, 1.8× fewer tool calls, the same culprit on all 20. The memory also stores its own failures: a plan that was right 2 times out of 4 was retired, with the tool weakness that misled it recorded in the graph.