CyberTriage AI

CyberTriage takes a cyber-fraud complaint and does three things: names the type of fraud, scores its urgency from 0 to 100 while flagging cases still inside a 48-hour "golden hour" window, and routes it to the right authority, such as a cyber cell, a bank's nodal officer or the Economic Offences Wing. It was built twice in 2026, for two hackathons. No result is on record for either.

First, in January 2026, a Model Context Protocol server: one Python file of 1,287 lines on FastMCP over HTTP, exposing 14 tools, 3 resources and 3 prompts, so that any tool-calling model can classify, score and route a complaint. The domain lives in configuration, not code. YAML and JSON files define the 13 cyber-fraud categories, and banking and IT-support configurations can be swapped in. Cases are kept in memory or in SQLite, and there are client demos for Gemini, OpenAI and Claude tool calling. Then, in April 2026, an OpenEnv reinforcement-learning environment of 512 lines: reset, step and state, three graded tasks with partial credit (classify, triage, and the full pipeline), 60 sample complaints, and a baseline agent that calls a hosted model.

Classification is keyword rules, not a trained model. No run was saved: there are no scores for the baseline agent, no throughput figure and no test suite, so this page counts what exists and nothing more. No agent was trained in the environment. Evidence extraction, multilingual support and links to police and bank systems are plans in the original proposal, not code.