Machine learning · 2025
Workflow Intake
Turn long document reviews into a trackable workflow.
A document-review workflow needs to show progress, recover from an interrupted session and make model-filled content recognisable. A background worker retrieves relevant material and extracts fields, then returns documents or spreadsheets with context for the reviewer.
- Status
- Production
- Domain
- Applied LLM · Cloud architecture
- Role
- Architecture, full stack
- Year
- 2025
Explore the workflow with illustrative content.
Keep the original input
The first stage represents an incoming record. A separate source copy makes it possible to inspect where an extracted value began.
Check the structured result
Fields are extracted into a consistent shape. Switch to the review example to see a missing field routed for attention.
Send it to the right next step
A complete sample reaches the output. A sample with a missing field remains in review rather than silently passing through.
Illustrative content · no live data
Turn an incoming document into a reviewable draft
Stack
- Azure Functions
- Cognitive Search
- Azure OpenAI
- Queue + Table Storage
- Vanilla JS
A document-intake platform that processes lease and tender material in the background and produces structured outputs for a reviewer.
Keep the job moving in the background
Submissions enter a storage queue and a scheduled worker processes them.
Persisted job state and attempt counters support progress tracking and retries when an upstream service is temporarily unavailable. The interface can restore pending-job information after a refresh.
Give the reviewer the source context
Search retrieves relevant document sections before language-model extraction.
The tender workflow surfaces those sections alongside the output, helping a reviewer check whether the relevant material was available and whether the resulting interpretation is supported.
Make generated content easy to spot
Model-populated spreadsheet cells are visually distinguished from template defaults.
That cue shows where content came from; it does not establish that a field is correct. It helps a reviewer prioritise the extracted values while retaining responsibility for the final document.
Design for interrupted work
Attempt accounting and completion state help the worker manage retries and repeated messages.
The application uses a straightforward web interface and established Azure services, keeping the submission, processing and review stages distinct enough to diagnose when something goes wrong.
Pipeline
Submission to a reviewable output
Stage Can reject
-
01
Submit
Job enqueued
-
02
Timer worker
Background processing
-
03
Search
Retrieved sections
-
04
Extract
Field values
-
05
Retry check
Job state updated
Transient failures can be retried
-
06
Review output
Populated workbook
Model-filled cells identified
Behind the work
The skills
that came together.
Cloud workflows
- Queue and scheduled-worker architecture
- Durable job state and retry handling
- Authenticated web interfaces
Document analysis
- Retrieval over document collections
- Structured extraction into templates
- Source context and reviewer cues
Project highlights
-
Trackable background work
Job state persists through processing and retries
-
Review with context
Retrieved sections available beside the output
-
Generated fields identified
Visual cues distinguish extracted content from defaults