RegLens

RegLens reads a bank's DataHub graph to find what a new regulation touches, costs acting now vs deferring vs minimum compliance, and writes the decision back so the next team inherits it.

Comment

Inspiration

Every regulated bank faces the same bottleneck: finite change capacity and a constant stream of regulatory change. The hard question isn't "what does this regulation affect?"... it's "should we implement it now, defer it, or do the minimum, and what does each choice actually cost?" Today that decision is made, when trying to remember which systems connect to what. DataHub already knows which systems connect to what. We built the agent that uses it.

What it does

RegLens turns a regulatory change into a costed investment decision.

  1. Discovers impact — it reads the bank's live context graph through the DataHub MCP server and walks lineage from the affected data asset outward. For our fictional regulation RCS-2026 (an enhanced customer-risk- classification standard) it finds 9 downstream assets automatically: datasets, a risk-scoring ML model, pipelines, and the regulatory reports and capital dashboards at the end of the chain.
  2. Costs three paths — a transparent scenario engine models ACT NOW, DEFER, and MINIMUM COMPLIANCE. It weighs not just build cost but the cost of inaction (probability × supervisory consequence + remediation) and the opportunity cost of the strategic programme the work would displace. Every number carries its assumption and a confidence, so a human can challenge it.
  3. Recommends and asks — it produces an Impact & Decision Card with a recommendation, then stops for human approval.
  4. Writes the decision back — on approval it attaches the assessment (a glossary term + a description) to every affected asset in DataHub. The next person or agent to open that asset inherits the whole decision instead of rediscovering it.

How we built it

  • DataHub as the source of truth. We seed a small, realistic fictional bank ("Northstar Bank", ~22 assets across Retail + Corporate) into a local DataHub quickstart, with full lineage from customer data through the risk-scoring model to capital reporting.
  • Read via the MCP server. The agent interrogates the graph through the official DataHub MCP server (search + lineage), so it discovers the blast radius rather than being told it.
  • Contribute back to the graph. The write-back uses DataHub's mutation surface to enrich the affected assets — this is the part that makes it an agent that does real work, not just a read-only analyzer.
  • A glass-box scenario engine. Deliberately simple and fully transparent: every figure is a named assumption you can edit, not a black box.

Challenges we ran into

Mapping a regulation to a data footprint is only tractable because the lineage exists in DataHub — the interesting work was turning that footprint into a defensible cost comparison, especially pricing inaction, which is inherently probabilistic. We kept the model honest by surfacing assumptions and confidence everywhere instead of hiding them behind a single number.

Accomplishments we're proud of

An end-to-end loop — discover, cost, approve, write back — where the write-back genuinely changes the graph, plus a recommendation that's fully auditable down to each assumption.

What we learned

The MCP server makes DataHub's context directly usable by an agent, and the highest-value move is closing the loop by writing knowledge back so it compounds.

What's next

Persist the assessment as first-class structured properties (not just a description), promote the reports and model to real Dashboard/MLModel entities, infer the anchor asset from regulation text with an LLM, and support a portfolio view across multiple concurrent regulations.

Note: RCS-2026 and its issuer are fictional by design, so RegLens makes no claim about any real-world regulation.

Built With

Share this project:

Updates