[02]
One agent per document type
A classifier routes contracts, filings and correspondence to dedicated extraction agents. Each path has its own tools and validation schema for the fields that document type should contain.
[CASE STUDY 03]
Golegal AI Legal Workspace
A consumer legal platform for drafting, review, signing, obligation tracking and access to regulated solicitors. I built five modules end to end.
- Role
- Full-Stack & AI Engineer — five modules, end to end
- Timeline
- Dec 2024 – Jan 2026
- Status
- Live
- Stack
- Nest.js, Next.js, PostgreSQL +4
[AT A GLANCE]
- OWNED
- Five modules end to end: document creation and RAG · e-signature and obligation extraction · legal pathways · tasks, calendar and notes · the services marketplace
- ALSO BUILT
- Payments and the token-metered usage module the paid plans run on
- NOT MINE
- The legal content and the regulated advice — Go Legal Ai is not a law firm, and the services are delivered by independent regulated firms
- PRODUCT
- Free · Power Up £19/mo · Pro £49/mo, metered in AI credits per action
- REACH
- England & Wales for instructed services; document generation across 150+ jurisdictions
[STACK]
- Nest.js
- Next.js
- PostgreSQL
- OpenAI
- Agentic RAG
- React Flow
- Plate.js
[01] The problem
Generic chat can’t be trusted with legal text
Legal teams lose hours to reading, extracting, and reformatting the same document types. Generic LLM chat wasn’t accurate enough to trust with legal text — it hallucinated citations and lost document context.
[02] The constraints
Why a better prompt was never the answer
Plausible is not good enough
A fluent but incorrect clause can create legal risk. The system had to ground claims in the user’s document and jurisdiction, making retrieval quality more important than a larger prompt.
Every AI action costs real money
Drafting, review and extraction cost per call and are sold as credits. Accuracy improvements had to justify their effect on unit economics.
The legal steps cannot be skipped, only made lighter
Drafting, review, signing and obligation tracking have real dependencies. The interface could simplify the sequence but not remove its controls.
[03] The approach
How it was built
The solution combined structure-preserving retrieval with document-specific extraction agents. Around that pipeline, I added credit metering and a single working surface for a legal flow a non-lawyer could complete.
[KEY DECISIONS]
- [01]Chunked documents by clause hierarchy so retrieved passages retain their context.
- [02]Routed document types to dedicated agents with focused tools and validation schemas.
- [03]Modelled legal matters as pathway graphs of drafts, AI steps, services and court forms.
- [04]Extracted dated obligations from signed documents into the user’s calendar.
* RETRIEVAL THAT KEEPS THE DOCUMENT SHAPE
[04] The trade-offs
What accuracy cost
[01]
Sell AI work as credits, priced per action
- Instead of
- One flat unlimited subscription, the way most SaaS is sold
- What it cost
- Customers must understand credits, and plan pages must explain each action’s price.
- What it bought
- Usage stays tied to model cost, protecting margins without making light users subsidise heavy ones.
[02]
Tune retrieval for precision, not coverage
- Instead of
- Retrieve generously and let the model decide what is relevant
- What it cost
- Chunking and routing require more work, and an occasional relevant passage may not surface.
- What it bought
- Fewer confident but misleading answers, the more dangerous failure in legal text.
[03]
One working surface for drafting, review and signing
- Instead of
- A wizard with a page per step, which is easier to build and easier to reason about
- What it cost
- One screen carries more state and requires stricter component boundaries.
- What it bought
- Users remain in the document instead of crossing multiple abandonment points.
[05] The build
The five modules
[01]
Retrieval that keeps the document’s shape
Documents are chunked by section, clause and sub-clause, with ancestry attached to every chunk. Retrieved passages keep their meaning and no longer require the model to infer missing context.
[03]
Legal pathways, as a graph you can see
Pathways models each legal matter as ordered actions on a React Flow canvas, from drafting and waiting periods to AI steps, services and court forms. Users can see the current step and its dependencies.
[04]
Signature, and what the signature commits you to
After multi-party signing, the platform extracts notice periods, renewal windows and payment deadlines from the executed document so commitments remain actionable.
[05]
The deadlines the documents created
Extracted obligations become calendar entries with tasks, notes and reminders. The dashboard leads with overdue and upcoming work.
[06]
When the AI is not enough, order a person
A priced services catalogue lets users order regulated legal work when AI is insufficient. Each service has a fixed fee plus VAT, jurisdiction is explicit, and the platform clearly states that it is not a law firm.
[06] The result
What Golegal measured
95%
Extraction accuracy
* on the evaluated set, Jan 2026
100+
Files / month
* in production, at handover
-30%
Manual document work
* reported by users, Jan 2026
50+
Active users
* legal professionals, Jan 2026
[HOW THESE WERE MEASURED]
The 95% is exact-match accuracy on the evaluated set, not every accepted document type, and that set was too small; the reflection below covers why. The −30% is user-reported rather than instrumented. File and active-user counts are platform records from January 2026.
[07] What I’d do differently
Honestly
I’d build the evaluation harness before the pipeline. Tuning by feel delayed clear decisions until a scored regression set existed.
Written by Faizan Amir
Role on Golegal: Full-Stack & AI Engineer — five modules, end to end
Last reviewed
Need something like this built?
Next case study [04]
Alfa
Multi-Role Fintech Platform
Read the case study →
