ChoiceCapture by ChoiceOS

AI-powered live-commerce capture that verifies sales, matches real products behind generic listings like “Item on Screen 147,” and produces live KPIs, evidence and fulfilment-ready records.

Comment

Inspiration

Live-commerce sellers can run hundreds of products during a single event, often using generic listings such as “Item on Screen 147” rather than the real product name.

This creates a major operational problem. After the live event, the seller may need to determine:

  • What product was actually shown
  • Whether it sold or did not sell
  • Who purchased it
  • The final sale price
  • Which inventory record should be updated
  • Where the item is stored
  • Which purchases belong to the same customer
  • How the event performed commercially

Existing workflows often rely on handwritten notes, spreadsheets, screenshots, platform order exports and manual reconciliation. They are slow, error-prone and especially difficult when auctions last only a few seconds.

ChoiceCapture was inspired by a real operational need within Choice Collectibles Ltd and forms part of the wider ChoiceOS commerce operating system.

What it does

ChoiceCapture is an AI-assisted live-commerce capture and reconciliation system.

It is designed to observe a live-selling event and create a dependable record of every product run, including sold items, unsold items and Buy It Now purchases.

The Build Week version focuses on a complete working flow:

  1. Start a live-commerce event in ChoiceCapture
  2. Observe auction and Buy It Now activity
  3. Detect when a product is run
  4. Determine whether it sold or remained unsold
  5. Capture the buyer, price, timing and available product evidence
  6. Match generic listings to real inventory products using GPT-5.6
  7. Prevent duplicate or false sale records
  8. Update live event statistics
  9. Provide live KPI data for an OBS browser overlay
  10. Produce post-event sales, product-run, evidence and fulfilment files

ChoiceCapture separates the platform listing title from the actual product identity.

For example:

  • Platform listing: Item on Screen 147
  • Observed product: Collectible Figure Example
  • Matched inventory item: CHO-FUN-0147

When a confident match cannot be made, ChoiceCapture preserves the original evidence and creates an unknown-product review record instead of silently guessing.

Why AI is needed

The transactional information and the real product identity are often disconnected.

A platform may report that “Item on Screen 147” sold, but that title does not tell the seller what physical item must be pulled, packed or removed from inventory.

GPT-5.6 is used to evaluate available product evidence together with structured inventory candidates and return a controlled result containing:

  • The most likely inventory match
  • Match confidence
  • Supporting evidence
  • Alternative candidates
  • Whether human review is required

The system is designed so AI assists identification and reconciliation without silently changing authoritative inventory records.

How we are building it

ChoiceCapture uses a modular architecture so the Build Week prototype can operate independently while later connecting directly to ChoiceOS.

The system includes:

  • A Chromium browser extension for event observation
  • A local capture and event-processing service
  • A product-run and sale state machine
  • Structured event and evidence storage
  • GPT-5.6 product matching
  • A live event dashboard
  • An OBS-compatible browser-source overlay
  • CSV and evidence exports
  • A post-live reconciliation workflow

The capture approach prioritises:

  1. Structured platform or page events
  2. DOM-based verification
  3. Visual product evidence
  4. OCR only as a fallback

This avoids depending on fixed screen coordinates, browser zoom or the seller keeping one part of the page visible.

Key technical challenges

Preventing false sales

Live interfaces can briefly display old or incomplete information while the next product loads. ChoiceCapture therefore does not create a confirmed sale merely because a price or buyer appears.

It uses a state machine that tracks:

  • Auction started
  • Auction running
  • Possible result
  • Auction ended
  • Result stability
  • Next auction started
  • Sale confirmed

Supporting very different auction lengths

Business users may run auctions lasting anywhere from one second to 30 minutes. The system therefore uses event-driven transitions and precise timestamps rather than relying on slow polling or a fixed confirmation delay.

Handling Buy It Now products

Buy It Now purchases can occur while a different auction is running and one listing can produce multiple purchases. ChoiceCapture models auction runs and Buy It Now purchases separately.

Identifying generic listings

The visible listing title may not contain the actual product name. ChoiceCapture combines visual evidence, prepared event data, inventory candidates and AI-assisted matching to connect the transaction to the real item.

Preserving trust

Every captured sale retains its original source information, evidence, timestamp and match status. Low-confidence cases are sent for review rather than being automatically treated as correct.

What makes it useful

ChoiceCapture is intended to turn a live event into operationally useful business records rather than only a sales total.

The resulting data can support:

  • Sales exports
  • Product-run analysis
  • Sold versus unsold reporting
  • Sell-through rates
  • Sales and revenue per minute
  • Viewer and buyer statistics
  • Product-performance analysis
  • Inventory reconciliation
  • Pull lists
  • Customer order bundling
  • Packing workflows
  • OBS live KPIs
  • Post-event evidence and audit records

What we are learning

This project is demonstrating that live-commerce capture is not simply a screen-reading problem.

Reliable results require a combination of:

  • Event-state modelling
  • Duplicate protection
  • Evidence preservation
  • Product identity resolution
  • Human review
  • Inventory reconciliation
  • Platform-specific adapters
  • Real-time and post-event analytics

We are also learning how to use GPT-5.6 as a controlled operational decision-support component rather than asking a model to make unsupported assumptions.

What is next

After validating the Build Week version, ChoiceCapture will be piloted within Choice Collectibles Ltd.

The next stages include:

  • Additional live-commerce platform adapters
  • Deeper ChoiceOS inventory integration
  • Automated stock reservations and deductions
  • Customer order consolidation
  • Pull, packing and dispatch workflows
  • Advanced viewer and product analytics
  • Configurable OBS overlays
  • Product-speed and pricing recommendations
  • Performance comparison across events, hosts and platforms

The long-term goal is for ChoiceCapture to become the live-commerce intelligence and operational capture layer within ChoiceOS.

Built With

Share this project:

Updates

posted an update —

A personal Build Week update: I only discovered OpenAI Build Week yesterday and decided to challenge myself to isolate ChoiceCapture from the wider ChoiceOS platform as a standalone MVP.

ChoiceCapture itself comes from a real operational problem within my live-commerce business, but preparing it as an independent submission in such a short timeframe has been an intense and rewarding experience. GPT-5.6 helped me structure the product, workflows and edge cases, while Codex helped me turn those decisions into working code and repeatedly test and refine the implementation.

I am proud of what has been achieved, particularly as this is my first software project and I have very limited traditional coding experience. Thank you for taking the time to view it.

Log in or sign up for Devpost to join the conversation.