01. Product & System Overview¶

What the product does¶

DRT AI Platform processes tribunal case PDFs into OCR markdown, structured extraction (entities / attributes / relationships), and UC1 automated factsheets that prefill application fields with provenance.

It is a private monorepo (backend/ + frontend/ + infra/) aimed at Debt Recovery Tribunal document workflows, with additional use cases described in plans but not yet coded.

Primary users and workflows¶

User (intended)	Workflow in this checkout
Registry / operator	Upload PDF → watch job stages → inspect OCR and extraction
Case worker	Select documents → generate factsheet → review fields + provenance
Admin / developer	Re-run OCR/extraction/merge stages; delete documents/factsheets

There is no login or RBAC UI. Access control is effectively “whoever can reach the API.”

Use-case status¶

ID	Name	Status
UC1	Automated Fact Sheet	Implemented (`uc1_factsheet/`)
UC2	Scrutiny Automation	Planned (`plans/phase_5_uc2_scrutiny.md`)
UC3	AI Search (orders/judgments)	Planned (`plans/phase_7_uc3_search.md`)
UC4 / UC5	Predictive analytics / STT	Out of scope for Phase 1 (PRD)

Major subsystems¶

Subsystem	Responsibility
FastAPI	Ingestion, job control, factsheet CRUD/status
Celery workers	Chunked OCR, extraction, merge, factsheet tasks
PostgreSQL	Documents, jobs, stages, chunks, artifacts, factsheets
MinIO	Raw PDF object storage
LiteLLM	VLM OCR, extraction LLM, factsheet LLM
OCR adapters	`vlm`, `sarvam`, `layout` (Paddle OCR service)
React console	Upload, jobs, documents, factsheets

High-level architecture¶

graph TD
  classDef default fill:#1e293b,stroke:#38bdf8,stroke-width:2px,color:#f8fafc
  classDef highlight fill:#065f46,stroke:#34d399,stroke-width:2px,color:#f0fdf4

  Upload["POST /documents"]:::highlight
  Job["DocumentJob + stages/chunks"]
  OCR["OCR stage"]
  Ext["EXTRACTION stage"]
  Merge["MERGE stage"]
  Art["Artifacts JSON / MD"]
  FS["POST /factsheet/generate"]:::highlight
  P1["Phase1 discriminators"]
  P2["Phase2 sub-models"]
  P3["Phase3 assemble + provenance"]

  Upload --> Job
  Job --> OCR --> Ext --> Merge --> Art
  Art --> FS --> P1 --> P2 --> P3

  linkStyle default stroke:#64748b,stroke-width:2px

Shared document pipeline feeds UC1. Search/vector services are optional infra only.

Request-to-result: document processing¶

%%{init: {
  "theme": "base",
  "themeVariables": {
    "primaryColor": "#1e293b",
    "primaryTextColor": "#f8fafc",
    "primaryBorderColor": "#38bdf8",
    "lineColor": "#64748b",
    "actorBackground": "#1e293b",
    "actorBorder": "#38bdf8",
    "actorTextColor": "#f8fafc",
    "labelBoxBorderColor": "#38bdf8",
    "labelBoxBkgColor": "#1e293b",
    "labelTextColor": "#f8fafc",
    "noteBorderColor": "#fbbf24",
    "noteBkgColor": "#78350f",
    "noteTextColor": "#fffbeb"
  }
}}%%
sequenceDiagram
  actor User
  participant UI as React
  participant API as FastAPI
  participant S3 as MinIO
  participant W as Celery
  participant LLM as LiteLLM / OCR

  User->>UI: Upload PDF
  UI->>API: POST /api/v1/documents
  API->>S3: Store object
  API->>W: Enqueue process_document_task
  API-->>UI: Document + job accepted
  W->>LLM: OCR chunks
  W->>LLM: Extract chunks
  W->>W: Merge
  UI->>API: Poll document/job status
  API-->>UI: COMPLETED + artifacts

Request-to-result: factsheet¶

%%{init: {
  "theme": "base",
  "themeVariables": {
    "primaryColor": "#1e293b",
    "primaryTextColor": "#f8fafc",
    "primaryBorderColor": "#38bdf8",
    "lineColor": "#64748b",
    "actorBackground": "#1e293b",
    "actorBorder": "#38bdf8",
    "actorTextColor": "#f8fafc",
    "noteBorderColor": "#fbbf24",
    "noteBkgColor": "#78350f",
    "noteTextColor": "#fffbeb"
  }
}}%%
sequenceDiagram
  actor User
  participant API as FastAPI
  participant W as Celery
  participant DB as Postgres
  participant LLM as LiteLLM

  User->>API: POST /factsheet/generate
  API->>DB: Create Factsheet PENDING
  API->>W: Enqueue factsheet task
  API-->>User: 202 + job_id
  W->>LLM: Phase1 discriminators
  W->>LLM: Phase2 sub-model fields
  W->>W: Phase3 assemble + provenance
  W->>DB: Store result COMPLETED
  User->>API: GET factsheet status / detail

Key technical bets¶

Bet	Tradeoff
Stage/chunk job model in Postgres	More schema complexity; enables resume and partial failure
Pluggable OCR engines behind one pipeline	Uneven metadata richness across engines
Structured LLM extraction + factsheet phases	Prompt/schema maintenance; LLM cost/latency
Docs/plans ahead of search & auth	Onboarding clarity vs code/docs drift

Interview answer: "What did you build?"¶

"A DRT document platform: upload PDFs, run a resumable OCR and structured-extraction pipeline on Celery, then generate multi-document factsheets with field-level provenance. UC1 is live; scrutiny and search are designed but not wired yet. The interesting part is treating jobs as stage/chunk state machines and driving factsheet fields from discriminators instead of one giant prompt."