The corpus was not reliably searchable
Arabic OCR, bidirectional text, image-based tables, and diagram-carried constraints made simple PDF search incomplete.
FIELD NOTES / REGULATORY INTELLIGENCE & GOVERNANCE
Regulatory decisions.
Connected to evidence.
Regulyn brings bilingual regulatory knowledge, spatial and policy analysis, formal review gates, and committee casework into one traceable workflow—so every answer and recommendation can be followed back to its source.
REGULATORY CASE STUDY
Regulyn
Bilingual document intelligence.
Governed committee casework.
Regulyn / Supervisory governancePRODUCT EXPERIENCE
ONE GOVERNED CASELOADEvery request has a state.
Analysis, endorsement, committee, recommendation, and decision progress remain visible together.
Follow the workflow ↗RegisterAnalyseValidateCommittee
DOMAIN
Regulatory & public-sector governance
WHAT WE DESIGNED
Document intelligence + case management
OUR ROLE
Product, AI & full-stack engineering
01 / THE INSTITUTIONAL CHALLENGE
A binding answer needs
more than confidence.
It needs the regulation, section, page, exact supporting text, authorized reviewer, case history, and the record of how a committee reached its recommendation.
Scanned Arabic and English regulations resist ordinary search. Diagrams and rasterised tables carry rules that text extraction misses. Meanwhile, development requests move through email, attachments, analysts, consulted entities, quality review, and committee meetings.
Regulyn was designed to connect regulatory knowledge and formal governance without allowing AI to become the decision-maker.
An answer without provenance was unusable
Analysts needed the exact clause on the exact page—not a fluent summary detached from its evidence.
Governance lived across disconnected tools
Completeness, analysis, endorsements, quality review, committee action, and minutes lacked one attributable trail.
02 / INSIDE THE PRODUCT
From incoming request.
To defensible committee action.
Explore the supplied product views behind caseload control, spatial evidence, quality gates, and regional retrieval.
A caseload with visible governance state.
Requests can be filtered by status, region, and type while the progress of analysis, endorsements, committee review, recommendation, and decision remains legible.
THE GOVERNED CASE LIFECYCLE
Register→Completeness→Analyse→Endorse→Quality gate→Committee→Minutes
03 / AI WITH EVIDENCE AND AUTHORITY BOUNDARIES
Let AI prepare the evidence.
Keep judgment human.
Regulyn separates extraction, validation, retrieval, analysis, and formal authority. Knowledge does not become queryable before review, and technical analysis does not pre-empt the committee.
Look under the hood ↘01
Extract—but do not silently publish
OCR and AI create a draft knowledge record for subject-matter review.
Index for cited retrieval
Approved clauses, figures, tables, page geometry, and access rules enter the searchable corpus.
Hold for human resolution
Original extraction and human corrections remain distinguishable and auditable.
02
Answer with page-level provenance
Document · section · page · text snippet · highlighted source region
03
Enforce the governance gates
The system informs and records; authorized people review, recommend, and decide.
04 / THE TECHNOLOGY & THE REASON
Retrieval with proof.
Workflow with control.
The architecture separates the bilingual product experience, versioned API, asynchronous document pipelines, hybrid retrieval, governed operational data, and access enforcement.
BILINGUAL EXPERIENCE
Next.js + React + TypeScript
Role-specific workspaces support English and Arabic, direction-aware layouts, streaming responses, and formal casework.
Why this combination?
One typed interface can serve analysts, reviewers, administrators, committee users, and external responders without flattening their roles.
VERSIONED SERVICE LAYER
FastAPI + Pydantic
Typed APIs keep intelligence, governance, export, and administration rules outside the presentation layer.
Why a strict API boundary?
Stable contracts, machine-readable errors, and centralized authorization are essential across a large operational surface.
OPERATIONAL & VECTOR DATA
PostgreSQL + pgvector + Qdrant
Relational records preserve cases, approvals, and audits while dense and sparse vectors support hybrid retrieval.
Why both relational and vector stores?
Governance needs transactions and history; evidence retrieval needs semantic, keyword, scope, and metadata filtering at query time.
DURABLE PIPELINES
Celery + Redis
Parsing, extraction, indexing, workflows, exports, and ad-hoc work run outside request-response paths with progress events.
Why asynchronous stages?
Long document jobs can resume from tracked, idempotent stages instead of restarting after a worker or provider failure.
SOURCE PRESERVATION
Commercial OCR + S3-compatible storage
Scanned text, tables, figures, annotations, page dimensions, and detected images are retained as source-linked artifacts.
Why preserve page geometry?
A citation must open the correct page and highlight the exact supporting region to be defensible.
ACCESS & AUDIT
JWT + Argon2id + document ACLs
Role checks, closed governance permissions, document-level access, rate limiting, and immutable audit records protect the decision trail.
Why filter access during retrieval?
Restricted content must never become a search hit, model context, or citation—not merely disappear from the final screen.
Preserve the source
Document, page geometry, tables, figures, and language metadata.
Approve the knowledge
Original extraction and reviewer correction remain separately attributable.
Enforce scope in search
Hybrid retrieval applies document and user access before evidence reaches the model.
Show the proof
Every supported claim can open against its precise source location.
The page uses the client-facing portion of the supplied case study. Product screens contain illustrative prototype data; displayed counts, places, statuses, and spatial findings are not presented as measured outcomes or verified client results.
Explore the complete documented stack
The documented implementation covers bilingual presentation, APIs, asynchronous processing, operational and vector data, object storage, identity, and retrieval quality.
- Frontend
- Next.js App Router · React · TypeScript · TanStack Query · Zustand · server-sent events · Arabic/English i18n
- API
- FastAPI · Pydantic v2 · Uvicorn · versioned endpoints · RFC 7807 errors
- Data & retrieval
- PostgreSQL · pgvector · Qdrant dense + sparse vectors · reciprocal-rank fusion · reranking
- Async & cache
- Celery · Redis broker/results · rate limits · session revocation · progress streaming
- Documents & AI
- S3-compatible storage · commercial OCR · multi-provider model resolution · human review manifests
- Security & governance
- asymmetric JWT · Argon2id · document ACLs · segregation of duties · immutable audit schema · malware-scan pathway
Technology names identify the documented implementation, not partnerships, certifications, or endorsements.
A CONVERSATION ABOUT YOUR REGULATORY WORKFLOW
Which decision still takes
days to evidence?
Bring one document-heavy regulatory, compliance, or committee workflow. In 20 minutes, we’ll map the evidence trail, human authority boundary, and a practical automation opportunity.
20 minutes · Your process, evidence, and next step
20 minutes · Your process, evidence, and next step