What's New in v2.5.2: Smarter OCR, Background Jobs & UI Polish

This release is the biggest improvement to our core processing engine since launch. We've significantly raised the accuracy bar for OCR extraction, added a proper async background job system so large files no longer time out, and shipped a visual refresh across the landing page and UI. Here's the full breakdown.

🔍 OCR & Processing Improvements

This is the headline update. If you've ever uploaded a multi-page statement and got back garbled numbers, missing rows, or incomplete transaction lists — this release directly addresses those problems.

Significantly Improved OCR Accuracy

We rebuilt key parts of the extraction pipeline around Mistral OCR. The new approach handles scanned documents, low-contrast images, and statements with non-standard layouts far better than before. In internal testing across a broad set of real bank statement formats, we saw meaningful reductions in extraction errors — especially around currency symbols, date formats, and decimal separators that previously caused mis-parsed transactions.

The improvement is most noticeable on:

Scanned PDFs (not digitally generated)
Statements with multi-column layouts
Documents with small fonts or compressed line spacing
International formats with non-US number conventions

Background Job System for Large Files

Previously, large or complex files would hit the request timeout and fail silently. That's now gone. We've introduced a full async background job system built on Inngest that handles files up to 50MB without any timeout risk.

Here's how it works: when you upload a file that's likely to take more than 30 seconds to process, the system automatically queues it as a background job and returns a job ID immediately. You can close the browser tab, come back later, and your results will be waiting. The conversion page polls for status and shows you live progress as the job moves through each step.

Each job runs through four clearly defined steps:

Download — file retrieved from secure R2 storage
OCR — Mistral processes the document (up to ~7 minutes for complex files)
Parse — transaction data extracted, categorized, and structured
Export — Excel, CSV, JSON, and QBO files generated and stored

If any step fails, the system retries automatically before surfacing an error. You can see the full execution trace in real time.

Automatic Retries and Better Error Handling

Failed conversions used to disappear into a generic error state with no explanation. Now each job has structured error states — if OCR fails partway through, the system retries the affected step, and if it ultimately can't recover, you get a clear error message explaining what went wrong rather than a spinner that never resolves.

Faster Processing for Multi-Page and High-Volume Uploads

Processing time for multi-page statements has improved. The async pipeline processes steps in parallel where possible, and we've removed several unnecessary round-trips that were adding latency between the OCR and parse steps.

✨ UI & Design Updates

Unified Design System

Every button, card, input, and label now uses the same consistent color tokens and spacing. The main visible improvement is button hover states — particularly the ghost-style "Sign In" button which was invisible on hover in dark mode.

Refreshed Landing Page

The How It Works section, feature cards, FAQ accordion, and pricing cards all received polish: consistent border radius, visible dark mode borders, smoother hover transitions, and a rebuilt Subscribe section with Turnstile spam protection.

What's Coming Next

Webhook notifications — get pinged when a background conversion finishes
Batch download — download all conversion outputs in a single ZIP
More export formats — OFX and MT940 on the roadmap

Have feedback or spotted an issue? Contact us — we read every message.