Product engineering / Demo-ready MVP

ClearStep.

Less paperwork. More clarity.

INTERACTIVE ILLUSTRATION / FICTIONAL NOTICE

NORTHBRIDGE / SAMPLE NOTICE

A deadline buried in the details.

Please submit your renewal form by 20 November. Include proof of address. No payment is required for this sample notice.

YOUR NEXT STEPS

From dense text to a clear next step.

Try the sample. No upload, model call, or personal data.

A browser-local explanation of the workflow, not the running ClearStep application.

WHAT IT DOES

A document assistant that turns dense notices into clear summaries, cited facts, and actionable checklists.

MY CONTRIBUTION

Solo project: application design and implementation across the document workflow.

THE INTERESTING DECISION

PDF text extraction and image OCR happen in the browser. Only extracted text is sent to the application's API; the original file is not uploaded to the server.

START WITH THE EVIDENCE10 MB

File-size limit documented in the README.

Inspect the source ↗

BUILT WITH

Next.jsTypeScriptOCROllama

PROJECT TYPE

Demo-ready MVP

START HEREREADME & setup

01 / THE IDEA

Less paperwork. More clarity.

The problem

A letter can contain an important deadline, a fee, and a required action, all buried in dense language. The challenge is making that information usable without sending uploaded documents to an external service.

The approach

ClearStep extracts text in the browser, then produces validated summaries, cited facts, and checklists. A bundled fictional notice makes the experience easy to try; optional local inference uses Qwen through Ollama.

02 / THE SYSTEM AT A GLANCE

Follow the flow.

01Your document
02Browser extraction
03Validated analysis
04Action checklist

A simplified architecture overview. The repository documents the implementation details.

03 / THE ENGINEERING DECISIONS

Why it's built this way.

Extract locally

PDF text extraction and image OCR happen in the browser. Only extracted text is sent to the application's API; the original file is not uploaded to the server.

Validate the output

Zod checks the structured response. Invalid model output gets one repair attempt before the application falls back to explicitly labeled demo analysis.

Make the demo dependable

A deterministic sample flow works without a model or paid API. It is labeled as demo mode, with conservative extraction for unfamiliar documents.

04 / WHAT TO LOOK AT

Evidence over adjectives.

Upload boundary10 MB

File-size limit documented in the README.

Model setupOptional

The bundled sample runs without Ollama.

PersistenceOn device

One recent structured analysis is stored in browser localStorage.

Capabilities and result snapshots follow the project's current README, reviewed October 10, 2026. Read the source

05 / THE HONEST PART

What I'd improve next.

Current boundaries

Scanned PDFs do not yet have a page-to-image OCR path.
OCR currently supports English and downloads language data on first use.
The deterministic sample is richer than analysis of arbitrary documents.

The next iteration

Add scanned-PDF OCR, more languages, and document-specific evaluation suites.

Open my Software & AI CV

KEEP EXPLORING

CloudCrew

Give agents a reliable place to work.