LabBrief

Inspiration

A vague request is a poor starting point for a purchase. People ordering data-preparation support need to see what they requested, what they are buying, and what the system will deliver. LabBrief explores that boundary with a deliberately small intake-and-checkout prototype.

What it does

Enter a synthetic request and choose from three catalog deliverables: a manifest brief, metadata checklist, or handoff plan. A local classifier suggests a category or abstains, while the interface retains exact request quotations and displays checklist questions. The user reviews and explicitly approves the scope and fixed quote before checkout.

PayPal Sandbox is central to the delivery gate. The server creates the order and controls the price. Returning from checkout or asserting success in the browser does not unlock a download: the server checks provider state, verifies the completed order and exact amount, and binds access to the originating session. The resulting ZIP contains the approved intake, Sandbox receipt summary, README, and a hash manifest. It is an intake bundle, not completed laboratory work or scientific analysis.

AI implementation

The default model is supervised local ML, not a generative chatbot. Scikit-learn TF-IDF features and logistic regression classify requests into manifest, metadata, or handoff categories. Domain, similarity, score, and margin guards can abstain. Questions are predefined templates and quoted facts are preserved verbatim. The model neither sets prices nor approves purchases.

Training uses 36 newly authored synthetic examples. A tiny same-author benchmark observed eight correct in-domain suggestions and one abstention across nine examples, and eight out-of-domain abstentions. These are limited development observations, not independent validation or general accuracy estimates. The model card explains failure modes and uncalibrated scores.

Implementation and new work

This independent prototype was created on October 29, 2026 during the hackathon period. Python 3.12 serves a loopback web interface; NumPy and scikit-learn support local classification; the payment adapter uses PayPal Sandbox Orders v2. The default AI path needs no paid API or runtime model download. OpenAI coding assistants contributed to review and documentation. No commercial product implementation, private research engine, proprietary weights, or customer dataset was imported. The independent source is MIT-licensed. The submitted public source is commit 750991df97ee3404cead5d9bec31d66434e43bd8 at https://github.com/mbasyoni/labbrief .

Challenges and lessons

The important boundary was separating advisory AI output from purchase authority. The application keeps the quote server-controlled, requires scope approval, and independently checks payment before delivery. Review also addressed cross-session access, request origin/CSRF checks, provider-response validation, repeated capture handling, safe rendering, and removal of private setup/recovery surfaces from the public source.

Testing and demonstration evidence

The video contains actual public-candidate interface footage followed by explicitly labelled historical private-test evidence. The public-candidate footage demonstrates local intake classification, quote review, approval, and the genuine fail-closed response when Sandbox credentials are absent. It is not a live payment recording.

An earlier exact private build completed a genuine USD 5.00 Sandbox capture and authenticated ZIP delivery on macOS on October 29. That payment result and download checks are historical. The released public tree removes private interfaces and strengthens capture validation; it has not completed a fresh real-provider transaction. Its payment regression tests use explicit fake providers. The 71 independent public-candidate checks include 42 adversarial capture cases; these do not substitute for a new provider acceptance run. Private account identifiers, credentials, and raw receipts are not distributed.

Run and evaluate

The public repository includes full source, synthetic data, MIT license, setup instructions, model card, and tests. Follow README with Python 3.12 and open http://127.0.0.1:8769 on the same computer. Local classification works without AI service credentials. The application is not publicly hosted.

Without PayPal Sandbox configuration, checkout remains disabled/fails closed and delivery stays locked. Sandbox checkout requires a developer Sandbox application and a separate Sandbox Personal buyer, configured securely as described in the README. No live-money flow or real card should be used.

Current scope and next steps

LabBrief is a loopback development prototype for synthetic requests. It does not perform scientific analysis or real-money sales. Session recovery, webhooks, refunds, taxes, multi-process deployment, broad platform acceptance, and production hosting are outside its current implementation. A fresh provider test of the exact public build and broader usability testing are the next validation steps, not completed claims.

Built With

• python
• scikit-learn