.png&w=256&q=75)
🤓 Latest Submissions
Actuarial Signal Radar
We detect first. AI investigates second. Humans decide. Traditional dashboards show what happened. Actuarial Signal Radar asks what was expected, how different reality was, and what an analyst should investigate next. Using official CNBS Honduras monthly motor-insurance data, the engine estimates prior-history trend and seasonality. Observed minus expected produces a residual. A 12-month robust z-score classifies NORMAL, WATCH or ALERT; standard z-score, EWMA and CUSUM add context. Claims are monetary accounting amounts, not accident or claim counts. Around 100 monthly observations favor an auditable, reproducible statistical baseline over complex ML that could overfit. AI never creates or overrides the alert. Evolus orchestrates validated signal handoff, curated Knowledge Library context, AMD Token Factory inference with DeepSeek-V4-Flash, and structured-response validation. The investigation summarizes evidence, proposes unestablished hypotheses, identifies missing information and recommends checks. Every accepted result requires HUMAN_REVIEW_REQUIRED. September 2025 claims were approximately L185.4M versus L142.4M expected: +L42.9M (+30.15%), robust z 11.02, ALERT. March 2025 claims-to-premiums ratio was 63.91% versus 53.05% expected: +10.86 percentage points, robust z 8.61; raw NORMAL became adjusted ALERT. Production end-to-end validation is verified for the flagship claims case. A second case exposed an intermittent "Missing model execution envelope" validation failure; safeguards were not weakened to bypass it. Cloudflare/OpenNext serves the dashboard; Railway runs Python. This is a lagged retrospective backtest with a conservative two-month availability assumption, not validated strict point-in-time performance. Source revisions and exploratory materiality remain limitations. This is anomaly detection and investigation support, not forecasting, fraud detection or causal inference. Self-hosted ROCm/vLLM was not validated.
18 Oct 2026
