SynapseGuard
An incident diagnosis agent based on a graph database that uses context such as service call graphs and change logs for root cause localization, exposing parameterized tool interfaces through an MCP server.
Timeline (last 90 days)
Developer Mohit Sachin Shirke built SynapseGuard for the Graph Hacks competition, a diagnostic agent that starts from an alert and follows the service call chain to find the root cause.
On a self-built dataset of 500 synthetic incidents, SynapseGuard ranked the true root cause first in 95.2% of cases.
On the same synthetic dataset, the strongest tabular baseline that removes relations and keeps the same signals achieved 81.6% accuracy, with the graph structure bringing about a 14-percentage-point improvement.
The author's first naive approach of 'select the deepest failing service' achieved only 25% accuracy on the synthetic dataset.
SynapseGuard was tested on the public RCAEval benchmark, covering a total of 645 real fault-injection cases across three systems: Online Boutique, Sock Shop, and Train Ticket.
On RCAEval under default settings, the graph structure's ranking is on par with the strong baseline, with wins and losses across different systems, and lags by 4 percentage points on the held-out Train Ticket.
SynapseGuard exposes a set of hand-written parameterized tool catalog (such as blast_radius, who_to_page, path_between, recent_changes) through an MCP server, rather than letting the LLM freely write Cypher queries.
Tested with an 8B local model running on CPU, it correctly selected the tool 12 times out of 12 questions.
Real data caught a bug that synthetic data would never expose: the severity score saturates at 1.0, making services indistinguishable from one another.
All Facts (9)
Developer Mohit Sachin Shirke built SynapseGuard for the Graph Hacks competition, a diagnostic agent that starts from an alert and follows the service call chain to find the root cause.
50%UnverifiedOn a self-built dataset of 500 synthetic incidents, SynapseGuard ranked the true root cause first in 95.2% of cases.
50%UnverifiedOn the same synthetic dataset, the strongest tabular baseline that removes relations and keeps the same signals achieved 81.6% accuracy, with the graph structure bringing about a 14-percentage-point improvement.
50%UnverifiedThe author's first naive approach of 'select the deepest failing service' achieved only 25% accuracy on the synthetic dataset.
50%UnverifiedSynapseGuard was tested on the public RCAEval benchmark, covering a total of 645 real fault-injection cases across three systems: Online Boutique, Sock Shop, and Train Ticket.
50%UnverifiedOn RCAEval under default settings, the graph structure's ranking is on par with the strong baseline, with wins and losses across different systems, and lags by 4 percentage points on the held-out Train Ticket.
50%UnverifiedSynapseGuard exposes a set of hand-written parameterized tool catalog (such as blast_radius, who_to_page, path_between, recent_changes) through an MCP server, rather than letting the LLM freely write Cypher queries.
50%UnverifiedTested with an 8B local model running on CPU, it correctly selected the tool 12 times out of 12 questions.
50%UnverifiedReal data caught a bug that synthetic data would never expose: the severity score saturates at 1.0, making services indistinguishable from one another.
50%