← Work / CrimCaseAI

Case study 01C · Legal intelligence

Useful legal intelligence without losing provenance.

CrimCaseAI is a portfolio of case-management and evidence-intelligence systems designed around sensitive material, restricted environments, and answers that must remain traceable to source evidence.

  • In-browser ML
  • Docker
  • Local OCR
  • Transcription
  • Vector search
  • Entity graphs
  • Grounded AI
01 · Role
Principal Engineer
02 · Domain
Legal intelligence
03 · Boundary
Privileged case material

01 · Problem

The screen was not the hard part.

Legal teams need software that can make large bodies of case material easier to navigate without weakening privilege, provenance, or control of the underlying evidence.

That creates a different product standard for AI: retrieval must stay connected to source material, processing boundaries must be explicit, and a confident answer without support is a failure state.

02 · System

A portfolio built around evidence boundaries.

The work moves from case management to evidence intelligence while preserving a consistent emphasis on structured data, local control, and grounded output.

  1. 01

    Case-management product

    The live SaaS product organizes case work, while a generation-two rebuild uses schema-generated forms and in-browser machine learning to strengthen the application model.

  2. 02

    Evidence intelligence

    A separate platform serving paying law firms turns case material into a navigable evidence layer rather than an unstructured document pile.

  3. 03

    Aegis on-prem appliance

    Aegis packages multimodal ingestion, local OCR and transcription, vector search, entity graphs, and citation-gated chat into a Docker appliance controlled by the firm.

03 · What made it hard

The answer is only useful if the source survives it.

The architecture treats provenance, local processing, and citation as product behavior, not implementation details hidden behind the interface.

01

Local processing boundary

The on-prem model keeps privileged material on hardware the firm controls and makes the deployment boundary understandable to the people relying on it.

02

Multimodal provenance

OCR, transcription, retrieval, and entity relationships remain tied to ingested evidence so analysis can be traced back through the system.

03

Citation-gated output

AI chat must ground material claims in available evidence rather than turning model confidence into an unsupported product promise.

04 · Next

Have a difficult system to make coherent?

Email StevenView experience