Regulatory Radar — Daksh Randeria

Background

Compliance teams at wealth management firms face a constant stream of regulatory output from the SEC, DOL, and CFPB. Most of it doesn't matter. Some of it does. And a small amount — a coordinated enforcement push across multiple agencies — represents a material risk that can take weeks to surface through normal monitoring channels.

The problem isn't access to information. It's the triage layer: identifying which regulations warrant serious attention, and getting ahead of broader enforcement patterns before they become headlines.

How the agent works

Regulatory Radar is built as a four-node state machine using LangGraph. Each node handles a distinct reasoning step, and the agent carries a structured "state" — a running digital clipboard — as it moves through the graph.

Fetcher — pulls the regulation title, summary, and associated agencies from Supabase, where regulations are ingested daily via the Federal Register API.
Triage — an LLM (Llama-3.3-70b via Groq) scores the regulation on relevance to wealth management (0–10), assigns an impact level, extracts primary keywords, and runs regime shift detection.
Investigator (conditional) — if a regime shift is detected, the agent autonomously triggers a targeted Tavily web search for law firm client alerts and recent enforcement actions, adding context a keyword search can't provide.
Writer — synthesizes everything into a structured executive brief with a Red/Amber/Green risk rating, a two-sentence summary, and three to four priority action items written for a compliance officer audience.

The conditional branch is the key design choice: the agent doesn't deep-dive everything. It reserves that compute — and the associated cost — for situations where it genuinely matters.

The "Regime Shift" signal

A regime shift is detected when any of three conditions are met: three or more agencies act in coordination on a single rule, a single agency exceeds a velocity threshold (five or more regulations in a seven-day rolling window), or a regulation scores 8 or higher on relevance. The velocity check is a PostgreSQL query against the Supabase database, not an LLM call — keeping it fast and consistent.

Combining a quantitative signal (velocity counts) with qualitative analysis (LLM relevance scoring) was a deliberate two-factor design. Either alone produces too many false positives. Together, they surface the handful of situations that actually warrant a deep dive.

Product thinking

The Streamlit interface is built around a single interaction: select a regulation, click Analyze, watch the agent reason. As it runs, the app surfaces each internal log step in real time — "Calculating SEC velocity...", "Joint action detected: 3 agencies...", "Fetching enforcement context..." — so a compliance officer or a recruiter can see exactly how the decision was made.

That transparency was intentional. A black-box output from a compliance tool is useless; the reasoning trace is what builds trust. The design also includes a "Demo Mode" that feeds a simulated high-velocity dataset, forcing the investigator branch to fire so the full logic tree is visible even when real regime shifts are rare.