Halley. Trustworthy AI for actuarial work.
Most AI tools can generate an answer. The real question is whether you can trust how it was produced. Halley is the StrinGaze Actuarial AI Agent — real software running today, working through structured models and governed knowledge so teams move faster without giving up control.
Four agent capabilities are live in the platform now — Model Analysis, Help Manual, Model Checkup, and Smart Modeling. Each keeps the actuary in the loop: Halley drafts, suggests, and explains; you accept, reject, or refine. What's coming below extends Halley from answering questions into orchestrating decisions.
An AI agent with a real engine underneath.
Halley doesn’t just talk about your models. It operates them — through the same API/MCP surface your team uses. The compiler underneath makes Halley fast. The security layer makes Halley governed.
Halley inherits the same guarantees every StrinGaze module gets — compiled performance, enterprise security, your choice of deployment. You don’t adopt new infrastructure to use Halley.
Halley’s runs execute on the StrinGaze compiler — three-tier dynamic compilation, OmniGraph dependency scan, CPU cache affinity.
Authentication, authorization, and version control apply to every Halley action — same audit trail your team already trusts.
Halley runs wherever StrinGaze runs — public cloud, private cloud, container, or on-prem.
Halley is only possible because of how StrinGaze mirrors the business logic.
Most actuarial AI assistants are bolted onto flat, code-buried models. They can read text — they can’t reason about the model. Halley is different because the StrinGaze platform already gave the model the structure that AI needs.
An LLM scans tens of thousands of lines of duplicated code per framework. Context window blown before it reaches reserves.
Halley reads one Class Tree node and infers every inherited variant. The model fits in a few hundred lines of structured nodes.
Assumptions are encoded inside lookup formulas. The LLM can’t separate "the rate dropped" from "the lookup index changed."
Smart Table is a no-code, structured assumption store. Halley reads it as data, not code.
LLM-suggested edits propagate unpredictably. No inheritance means every product needs separate review.
Halley edits a parent node; inheritance auto-applies to children. One review, framework-wide effect.
Other actuarial AI assistants exist. But they’re trying to reason about flat, duplicated, code-buried models. Halley doesn’t have to — because the StrinGaze Class Tree and Smart Table already gave the model the structure that AI needs.
Running today. More on the roadmap.
Halley is real software, not a deck. Two capabilities you can see today, and a roadmap built on the same Class Tree and Smart Table foundation. Every step keeps the actuary in the driver’s seat.
What Halley does today.
- Model Analysis — Ask about any variable, table, formula, or Class Tree structure — Halley traces how inputs flow through the calculation chain and explains inheritance and overrides.
- Help Manual & Model Checkup — Step-by-step answers grounded in StrinGaze product guides; quality rules run on a selected model to return a health score, severity ranking, and ranked fixes to review.
- Smart Modeling — Plan, then apply model changes in your workspace — plan → confirm → execute, with the actuary in control at every step.
Where Halley is going.
- Phase 1 · Design-to-Model — Design documents become runnable models: graded structured parameters, mapped to a master product, run on a dual basis (pricing-basis premium + best-estimate profitability). The foundation.
- Phase 2 · Pricing & FP&A agent layers — Agents that orchestrate analysis, planning, and steering across the lifecycle, built on the same governed orchestration. Acceleration.
- Phase 3 · Governed Decision Intelligence — The Governance Agent and enterprise foresight turn outputs into trusted, forward-looking decisions. Leadership.
An AI you can put in front of your auditors.
Five principles, applied to every Halley action — not as a marketing layer, but as platform-level enforcement.
Our AI agent is split into a doer and a checker. One executes; the other reviews logic, structure, and result — and the actuary is always the final reviewer.
Our Class Tree structure makes table data and code ~80% less. It is AI-native and easy for AI to understand — Halley reasons about the model, not just the text.
Our knowledge base is built on governance to ensure consistent results no matter which LLM you run — and you can swap models freely.
Embedded best-practice structure means anyone who builds gets the same result — models stay simple, transparent, well-maintained, and far less prone to hallucination. Non-standard models accumulate inefficiency and need periodic optimisation.
Halley drafts, suggests, explains; actuaries accept, reject, refine — no autonomous submission. Every action is logged (prompt, model version, output, reviewer, timestamp). Defensible to your auditors.
Halley runs where StrinGaze runs — public cloud, private cloud, or on-prem. Your data never trains anyone else’s model.
Halley exists to make actuarial work more meaningful.
Judgment on assumptions. Design of new products. Interpretation of results in a board meeting. This is the work an actuary trained for — and it’s the first work that gets squeezed when cycles overrun.
Halley takes the mechanical load — wiring assumptions, validating tables, reconciling against incumbent platforms, porting legacy models — so your time goes back to the work only an actuary can do.