Reliable, Smart Workflows for Finance, Wealth and Insurance — Karthic Rao, Perpetual
Summary
This presentation outlines an architecture for building reliable and intelligent workflows in highly regulated sectors (Finance, Wealth, Insurance). The core challenge is reconciling the need for deterministic compliance (reliability) with the ambiguity of real-world processes (intelligence). The proposed solution is a hybrid system where control flow, evidence, and external actions are managed by a deterministic system, while Large Language Models (LLMs) are restricted to explicit, typed points for interpretation and judgment. This approach ensures that any judgment made by AI must pass through a compiler that validates it against predefined 'action contracts,' creating immutable, auditable plans.
Key takeaways
-
The Reliability vs. Intelligence Boundary
In regulated industries, reliability and intelligence are not contradictory. The key is defining the boundary: Determinism must govern the flow of control, evidence, and side effects, while intelligence handles interpretation and judgment (e.g., reading ambiguous instructions or client responses).
-
The Problem of Ambiguity in Wealth Management
3:39
A simple request, like 'Fill out the W-9 form for Mr. Harper,' can become a multi-day process involving resolving ambiguity (e.g., identifying the correct 'Harper' family), gathering missing data, and obtaining multiple signatures. Fixed flowcharts fail when ambiguity arises.
-
Deterministic Planning and Compilation
10:45
The system uses a planner to translate a request into a semantic graph. This graph is then passed to a compiler, which validates it against the firm's action contracts. The compiler returns a valid plan, a clarification request, or a refusal. Approved plans are immutable and hashed.
-
State Management and Controlled AI Usage
Waiting and follow-ups are managed as state machine steps, not automatic retries. External actions (like sending an email or requesting a signature) have separate gates tied to the exact payload, ensuring no action leaves the company without explicit approval.
-
Scaling Inference and Compliance
The design targets using smaller, 'Flash'-class models for the majority of decisions (e.g., 19 out of 20 decisions), reserving the larger, frontier model only for the most complex cases. This allows for private, locally hosted inference while maintaining compliance.
Technical details
-
Workflow Architecture
720s
The system separates concerns: the Planner translates the request into a semantic graph (nodes point back to the request). The Compiler validates this graph against the firm's action contracts (e.g., 42 procedural nodes covering models, documents, emails, signatures).
-
Action Contracts and Gates
Every external action (e.g., sending an email, requesting a signature) must pass through a dedicated gate. This gate requires explicit, typed payload approval, ensuring the action is safe and known to the company.
-
State Machines and Auditing
Waiting periods and follow-ups are modeled as state machine steps. The system tracks the provenance of data (how and when a value was obtained), which is critical for compliance. The process is designed to handle partial responses, pausing the workflow until confirmation.
-
Model Efficiency and Deployment
The architecture minimizes reliance on large models by designing workflows such that 95% of decisions can be handled by smaller, specialized models. This allows the system to achieve local, private inference, significantly lowering compliance risk.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.