How a Michigan county IT director built a production-grade AI legal research system — from a laptop proof of concept to FedRAMP High architecture on AWS GovCloud — in under sixty days.
Prosecutors working in county offices — unlike their counterparts at large law firms or state agencies — rarely have dedicated legal research staff. When a charging decision needs to be made, an assistant prosecutor must navigate hundreds of Michigan Compiled Laws sections, cross-reference sentencing guidelines, and verify that every element of a charge can be proven at trial. All of this happens under time pressure, often from an office without deep research infrastructure.
Consumer AI tools like ChatGPT offered the promise of faster research — but introduced a far more dangerous problem: hallucination. An AI system that confidently cites a statute that doesn't exist, or misquotes a charging element, creates Brady disclosure obligations, suppression risks, and in the worst case, wrongful prosecution. In a legal context, a plausible-sounding wrong answer is worse than no answer at all.
The Grand Traverse County IT team identified a precise requirement: any AI system used in a prosecutorial context must pin every answer to a verbatim, verified source. Not a paraphrase. Not a summary. The actual statutory text — cited, traceable, and auditable.
"Hallucination is not an acceptable risk in legal proceedings. Every answer must trace to a verbatim source."
BLACKSTONE Design Principle — Grand Traverse County ITRather than deploying an existing legal AI product — most of which are built for large law firm use cases at large law firm price points — the Grand Traverse County team chose to build a purpose-built system grounded in a single architectural constraint: retrieval-augmented generation with citation pinning.
In a citation-pinned RAG system, the AI model never answers from memory. Every response is generated by first retrieving the most relevant chunks of verified statutory text from an indexed knowledge base, and then instructing the model to answer only from those retrieved passages — with the source document and exact text included in every response. If the knowledge base doesn't contain an answer, the system says so.
This constraint eliminates hallucination at the architectural level. It also creates a natural audit trail: every answer BLACKSTONE provides is backed by a specific MCL section, a specific passage, and a timestamp — meeting Brady documentation requirements without additional workflow overhead.
The initial proof of concept was built in a single development sprint on a standard Windows laptop. Five Python files. A ChromaDB vector store. A Flask web interface. Four MCL documents ingested and chunked into 161 searchable passages. The entire stack ran locally, with no cloud dependency.
On April 7, 2026, the proof of concept was demonstrated live to Prosecutor Noelle and her team at the Grand Traverse County Prosecutor's Office. The demo covered live citation-pinned answers on PRV and OV sentencing scoring questions — the kind of detailed statutory lookup that previously required manual research. The prosecutors asked questions in real time. BLACKSTONE answered each one with the verbatim MCL passage it drew from.
The prosecutors approved moving to production the same day.
Moving from a local proof of concept to a production system used by a county prosecutor's office required a fundamental rethinking of the hosting environment. Legal data — statutes, charging documents, case research — cannot live in a standard commercial cloud environment. The production architecture for BLACKSTONE was designed around FedRAMP High compliance from the first infrastructure decision.
The three-tier access model — Admin, ReadOnly, and Auditor — was designed specifically for the prosecutorial workflow. Assistant prosecutors query the system through the ReadOnly tier. Compliance staff and defense discovery review use the Auditor tier, which provides full query logs and source citations for Brady documentation. No user can modify the knowledge base through the query interface.
The initial BLACKSTONE knowledge base focuses on Michigan Compiled Laws. Phase 2 extends the platform into the full workflow of a county prosecutor's office through three targeted integrations:
Extend the knowledge base beyond MCL statutes to include case law, annotations, and secondary sources through Westlaw's API. Citation pinning applies equally — every case citation traces to a verified Westlaw source record.
Connect BLACKSTONE to the county's existing case management system via NIEM web services, allowing charging guidance to be generated in the context of a specific case record — not just in isolation.
Ingest Axon-generated transcripts as searchable text documents. Prosecutors will be able to query BLACKSTONE against the factual record of an incident — "given what the officer observed, what charges apply?" — with every answer still traced to statutory authority.
Designing hallucination out of the architecture — rather than trying to detect it after the fact — was the foundational decision that made BLACKSTONE usable in a legal context. Every subsequent design choice flowed from that constraint.
FedRAMP High was a first-class requirement from the first infrastructure decision, not an afterthought applied during security review. This approach adds cost and complexity early — and eliminates far larger costs later.
The proof of concept was built to answer real questions from real prosecutors — not to demonstrate technology capability in the abstract. The specificity of the demo (PRV/OV sentencing scoring, live questions) is what earned same-day approval.
BLACKSTONE is not a general-purpose AI assistant. It is a purpose-built legal research instrument. Scope discipline — knowing exactly what the system is and is not for — is what makes it trustworthy in a high-stakes environment.
We built BLACKSTONE in Grand Traverse County, Michigan. The governance frameworks, architectural patterns, and lessons learned are available to peer counties ready to build responsibly.
Start a Conversation