EvidenceLens

Inspiration

A payment record, a seller’s delivery message, and a buyer’s complaint describe different things. Combining them into a confident narrative can hide the central disagreement. EvidenceLens explores a focused question: can an AI-assisted workspace help digital-service sellers and support operators prepare a draft while keeping its sources, contradictions, and missing evidence visible?

The prototype centers on a deliberately difficult example: the seller reports ten delivered files, while the buyer reports four present and six missing. The goal is a reviewable account of the evidence, not a prediction of who should win.

What it does

EvidenceLens is a single-case workspace for PayPal Sandbox dispute evidence. It organizes sources and a timeline, highlights conflicting delivery accounts, and connects draft statements to exact source passages. Reviewers can inspect citations beside the draft, edit statements, and mark each statement reviewed. Editing a statement resets its review status.

The default sample runs with deterministic local rules. Optional AI review suggests claims, contradictions, and gaps without overwriting the draft. Screenshot transcription is staged separately: reviewers compare the extracted text with the image, correct it, explicitly label simulated material, and admit the transcript. Original image bytes and transcription history stay linked in the JSON packet.

JSON and TXT exports preserve unresolved issues and review status. Nothing is filed with PayPal; the app cannot issue refunds, process payments, or access production funds.

How we built it

The application uses a dependency-free JavaScript frontend and a Worker-compatible backend. A deterministic engine handles import, analysis, citation integrity, and export. Server-side adapters implement PayPal Sandbox OAuth and read-only Disputes API retrieval.

The deployed AI configuration uses DeepSeek Flash for text review and image transcription, with reasoning effort set to none. Calls require an explicit action and consent; credentials remain server-side. OpenAI coding assistance was used during development, rather than as the deployed model provider.

Transient results are separated by per-page capabilities. There is no database or saved workspace; refreshing clears the working page. The public source is MIT-licensed and includes automated tests and a Python/JavaScript parity oracle.

Challenges we ran into

Exact quotations are useful but insufficient: a correctly quoted “four of ten” must not become a claim of complete delivery. We added citation and adverse-evidence guards while keeping semantic support explicitly unverified. Earlier model outputs failed validation; a later accepted output did not justify removing those guards.

Asynchronous calls introduced another risk: a delayed result could overwrite newer edits. Snapshot checks and explicit recovery protect the workspace without silently repeating provider calls. Image intake required its own admission step so machine transcription would not immediately become trusted evidence.

Accomplishments that we’re proud of

The hosted desktop workflow passed an end-to-end check: a real Sandbox read, four synthetic records, actual OCR of a synthetic PNG, explicitly reviewed and synthetic-labeled admission, actual AI review, citation inspection, review-state reset after editing, and both exports.

That run produced five AI claims, two contradictions, and three gaps. The exported packet contained six sources and six draft claims. Response-origin provenance was verified, but semantic support and completed review remained false. This was an automated workflow demonstration, not independent human validation.

The public source also passed 116 automated tests, 13 parity snapshots, and built-artifact checks.

What we learned

Useful review assistance needs visible boundaries between source content, model interpretation, and reviewer decisions. Preserving an unresolved contradiction can be more valuable than producing a smoother narrative. A successful integration call, an accepted model response, and an approved draft are separate milestones.

What’s next

Next steps are usability evaluation with intended users, broader browser testing, and stronger evaluation of unsupported conclusions and transcription errors. Production use would require additional privacy, security, and operational work. For now, EvidenceLens is a public prototype for test data only, not a system for real customer records or automatic dispute submission.

Built With

• deepseek
• javascript
• paypal