01 / Problem and response
The design problem was not simply finding documents.
Cross-country legal research requires analysts to locate relevant instruments, determine whether a source covers the right jurisdiction and period, interpret provisions against a structured question, and preserve enough evidence for another researcher to review the judgment.
A single-pass assistant would collapse those activities into one opaque response. I translated the existing research process into a staged system where automation could support selected steps without inheriting authority over final coding decisions.
Workflow architecture, orchestration and agent logic, evidence and confidence states, deterministic routing, analyst review outputs, and layered validation.
02 / Decision log
Three choices shaped the system.
Each choice trades some apparent simplicity for a workflow that can be inspected, tested, and handed back to a domain expert.
Separate retrieval from substantive judgment.
Finding a plausible source is not the same as establishing that it supports a legal conclusion. Discovery proposes candidates; later stages evaluate eligibility and support.
More orchestration and latency in exchange for clearer error boundaries and reviewable evidence.
Route evidence through explicit states.
Structured status fields and deterministic rules make it clear what may advance, what needs another pass, and what must be held for review.
More schema and control logic in exchange for predictable handoffs between probabilistic components.
Make human approval a system state.
The analyst receives traceable evidence and a defined decision point. Human review is part of the operating flow, not a disclaimer after generation.
The pilot supports rather than replaces expert work; the final throughput depends on review capacity.
03 / Validation
System behavior and substantive quality were treated as different questions.
Software checks covered deterministic behavior, state transitions, expected schemas, routing rules, and edge cases. Pilot outputs were then compared with manually coded reference material and prepared for analyst review.
Passing software checks establishes that the workflow behaves as specified. It does not, by itself, establish that a legal conclusion is correct.
- Functioning multi-stage controlled pilot
- Structured evidence and review artifacts
- Documented operating logic and handover package
- Organization-wide or recurring adoption
- Measured productivity or cost impact
- Fully automated legal accuracy
Time-saving estimates and planned integration were planning assumptions, so they are excluded from the measured results presented here.
04 / Public boundary
What can be shown publicly.
The case describes the problem, my role, the top-level control model, and the reasoning behind key design choices.
Internal code, prompts, datasets, reports, screenshots, and detailed test results remain outside the public portfolio.