Loading…
RAG Development · Chicago, IL
A commercial lines underwriter at a Loop carrier reviews 40 to 60 policy submissions a week. Each submission carries a base policy form, between six and twenty endorsements, prior loss runs, broker submission emails, and AM Best ratings on the primary and reinsurance carriers. The judgement calls are the interesting part. The document hunting is not.
A senior associate at a Chicago AmLaw 100 firm running document review on a $4M e-discovery matter is sitting in front of Relativity for ten hours a day. The questions are repetitive: does this email mention the 2022 supply agreement, does this attachment include privileged advice, does this contract contain a most-favored-nation clause.
We build RAG systems for Chicago insurance, trading, and legal e-discovery teams, with engagement scope and pricing set to fit the workflow.
Describe the document review workflow you want to compress.
Insurance underwriting at Aon, Allstate, or a Lake Shore Drive MGA pulls the same shape every week. A submission lands. The underwriter has to read through a base form (typically an ISO CG 00 01 or carrier manuscript), pull out the endorsements that materially change coverage, check the prior loss runs against the loss pick, validate that the reinsurance treaty actually covers the cession. Three of those four steps are document retrieval. The fourth is the actual judgement call.
Trading firms operating in the Chicago futures markets, from Citadel and Jump on the prop side to the in-house desks at CME Group, deal with exchange notices, contract specification changes, position limit advisories, and clearinghouse risk bulletins. When the CFTC issues a new position-limit rulemaking, somebody on the compliance team has to map it across every contract a trader is permitted to touch. Today that's a day. With a properly tuned retrieval system over the regulatory corpus, it is twenty minutes.
E-discovery is the third pattern, and Chicago has a uniquely concentrated stake because Relativity (the dominant review platform) was built here. A second-pass review on a 2 million document corpus at $0.18 to $0.35 per document for contract attorneys lands between $360,000 and $700,000 before partner review. A retrieval layer on top of the Relativity workspace that surfaces the 50,000 documents actually relevant to the privilege question takes that cost down by an order of magnitude.
The unifying observation: in every one of these workflows the expensive humans are doing keyword search and pattern matching. Generic enterprise search products (the kind Outcome Health or Enova might already have licensed) don't understand the semantics of an ISO endorsement form, a CME contract spec, or a privilege log. A RAG system built around the actual document shape does.
Six components, each shaped by the document realities of insurance, trading, and e-discovery work.
ISO standard forms, manuscript endorsements, declarations pages, loss runs in carrier-specific PDF layouts. We parse each form ID separately and preserve the endorsement-to-base-form override relationship so coverage queries return the actually-effective language.
Direct connection to RelativityOne or on-prem Relativity via the REST API. Document text plus existing coded fields (issue tags, privilege calls) feed the vector index. Citations link back to the document control number so reviewers stay in their existing tool.
CME Group contract specifications, position limit notices, clearinghouse advisories, and CFTC rulemakings. We track effective dates and supersession chains so a query for current rules doesn't return a 2019 advisory that's been replaced.
BM25 plus dense embeddings (text-embedding-3-large or Voyage for legal-specific corpora). Metadata pre-filters on line of business, jurisdiction, effective date, and document type keep result sets focused without sacrificing recall.
Cohere Rerank or BGE Reranker tuned against your real questions. Underwriting and privilege-log queries are exactly the use case where re-ranking earns its keep, raising top-5 precision by 10 to 15 percent versus a vanilla embedding search.
Every query and answer logged to immutable storage (S3 Object Lock or equivalent) in the structured format your archive vendor (Smarsh, Global Relay) already consumes. FINRA Rule 4511 and CFTC Regulation 1.31 retention windows configured at the bucket level.
Chicago's economy is heavier on document-driven work than the coastal stereotypes suggest. Aon, Allstate, CNA, and Zurich North America anchor the insurance side. Citadel, Jump Trading, DRW, and the CME Group itself anchor the derivatives and futures side. Morningstar runs out of Chicago. Relativity, the dominant e-discovery platform globally, is headquartered here. Outcome Health, Enova, and Sprout Social represent the tech end. Each of these operates on internal documents that nobody else has access to.
The regulatory texture in Chicago tilts toward insurance and futures. Illinois Department of Insurance Part 916 governs claim and underwriting data handling. The CFTC and the National Futures Association regulate any system touching futures positions, with Regulation 1.31 mandating five years of records on books and records (a category that arguably includes AI-generated answers about position limits). For e-discovery work, the relevant standard is FRCP Rule 26 proportionality, which an AI-accelerated review actually helps satisfy by reducing the hours required to produce defensible work product.
We work with Chicago firms remotely. Discovery calls run on central time. Deployments target whichever cloud you already operate in: AWS us-east-2 (Ohio) is the standard choice for proximity to Chicago, but Azure Central US and on-prem VPC-restricted deployments are both standard for compliance-heavy workloads.
Tell us which document workflow is eating the most hours. We reply within one business day with a rough scope and price range.