LiveRegTech · Capital-markets compliance

Adjudicator (SEBI)

Builder · Product & architecture · 2026 — ongoing

An agentic-RAG compliance assistant: given a product-change scenario it retrieves the currently-valid SEBI obligations, traces supersession to the consolidated position, and flags gaps — every finding cited back to a source circular.

The problem

For a given product-change scenario, which currently-valid SEBI obligations apply? Answering means tracing a circular through its amendments and master-circular consolidation to the live position. The worst failure is a false negative — a real obligation never retrieved, i.e. a silent compliance gap.

Approach

Recall-first by design: a missed obligation is the costliest error, so retrieval optimises for recall before precision.
Hybrid candidate generation (dense + BM25/FTS) then a Claude Haiku reranker — measured MRR 0.79 → 0.97 on a 20-query golden set.
Hand-rolled Anthropic SDK ReAct tool-loop (no framework) so routing, multi-hop, corrective retrieval and the grounding gate are all owned and explainable.
Temporal + supersession metadata and a narrow citation-graph tool let the agent follow amendment chains to the consolidated position.

What it does

Takes an unstructured scenario, decomposes it, and calls retrieval tools to gather candidate obligations.
Traces supersession and cross-references to reach the consolidated, currently-valid set.
Synthesises a grounded finding for each gap — every claim cited to a source circular, flagged ‘for expert review only.’

Outcome & results

Recall@3 1.00, precision@1 0.95 with rerank on the golden set (EXP-001).
Eval harness with a golden dataset and trajectory evaluation — right tools, right order, correction fires when it should.
Every choice decision-logged as an ADR: corpus boundary, embedding model, retrieval architecture, agent framework.

Claude (agent + Haiku rerank)

Anthropic SDK

FastAPI

Next.js

Supabase / Postgres

pgvector

BM25 / FTS

Live Code CurveIQ