AI underwriting & claims
you can actually audit.
Aegis is a hybrid rules-plus-LLM decision platform for insurance operators. Every decision is grounded in deterministic risk models, logged with full rationale, and works offline out-of-the-box.
POST /decide/underwrite
Auto policy · age 19 · high risk
Risk score
77/100
Confidence
72%
Suggested premium
$124.10
Model
gpt-4o-mini
Reasoning
- ▸Base rate for 'auto': 0.060
- ▸Policy type 'auto' classified as elevated risk (+10)
- ▸Young applicant (age 19) increases risk (+12)
- ▸Declared risk profile: HIGH (+20)
Platform
Everything you need to ship auditable AI.
Grounded Decisions
Every AI call starts from deterministic risk & fraud rules, so the LLM decides on top of facts — not hallucinations.
Auditable by Design
Every decision is persisted with inputs, reasoning, model, and confidence. Regulators love it.
Provider-Agnostic
Swap OpenAI, Anthropic, or a local Ollama model with one env var. Ships with a zero-config heuristic fallback.
Real Underwriting
Risk score, premium recommendation, and explainable rationale on every policy application.
Claim Fraud Triage
Fraud score, leverage analysis, out-of-window detection — and a FLAG / REJECT / APPROVE decision you can stand behind.
Streaming Copilot
An SSE chat endpoint gives operators a natural-language copilot grounded in the live decision log.
Architecture
Rules first. LLM second.
A deterministic risk engine computes the numbers and produces structured factors. That block is then passed to the LLM as grounded context. The model's only job is to pick a decision and summarize the rationale — which is then stored in an append-only audit log.
rule_engine
└─ risk_score, factors[], suggested_premium
│
▼
llm(system + structured_context)
└─ { decision, confidence, reasoning[] }
│
▼
Decision table ◀── audit log, immutable
│
▼
GET /decisions ─── compliance & dashboardsShip AI decisions your lawyers will sign off on.
Open the dashboard, file a claim, and watch the underwriter agent reason about it in real time. No credit card, no API key required.