ANT — Purchase intelligence you can verify

ANT is a mobile purchase-intelligence companion for the everyday spending people rarely remember clearly: a coffee between meetings, a Friday snack, a small top-up at the supermarket, one more thing in the basket. None of those moments seems important alone. Across weeks, they can quietly shape someone’s budget, shopping routine and confidence in the choices they are trying to make.

ANT turns confirmed receipt data into understandable, evidence-backed discoveries. It helps a person see where small purchases are accumulating, recognise recurring products before an unnecessary repeat shop, spot changes in the cost of their usual basket, and turn a pattern they care about into a practical goal. Instead of showing a vague score or demanding that users log every habit, it works from receipts they already have.

The problem ANT solves is not a lack of data—it is the gap between what people think they bought and what their receipts can actually show. A person may feel that they are spending carefully, buying fewer snacks, or planning healthier shops, yet still miss the recurring purchases hidden in a busy week. ANT makes those patterns visible early enough to support a more intentional next decision.

Every conclusion is explainable: ANT shows the period analysed, the receipts used, the calculation and the limitation. It only describes what was purchased; it never claims to know who consumed an item, whether it was shared, or what it means for someone’s health. The result is a gentle, useful mirror—not a judgement.

This is a runnable Android submission for OpenAI Build Week · Apps for Your Life.

Why ANT

Most spending tools show a total after the money is gone. Most wellbeing tools ask people to remember every detail of their habits. ANT starts with something people already have: receipts.

Each insight is designed to be inspectable. ANT shows the period analysed, supporting receipts, calculation and limitation. It can say what was purchased; it cannot know who consumed it, whether it was shared, or what it means for someone’s health.

Judge quick start

What	Details
Platform	Android 8.0+ (API 26+)
Test build	artifacts/hackathon/ANT-hackathon-demo.apk — upload this exact file to a public release link before submitting
Network	Not required for the judge route
Account / API key	Not required
Demo data	Synthetic only; no personal receipts or real merchant brands

Install the APK, then follow this route:

1. Complete or skip onboarding.
2. Open Scan and choose Use prepared demo receipt.
3. Select Process receipt.
4. In Review receipt, correct the highlighted uncertain product and confirm it.
5. Open See evidence to inspect the receipts, calculation and limitation.
6. Open a likely-needed prediction and add it to the shopping list.

The critical route is covered by Android instrumentation testing.

What is implemented

• Kotlin + Jetpack Compose Android application with a complete mobile flow.
• Editable receipt review, local quality checks, duplicate protection and Room persistence.
• Deterministic Kotlin calculations for discoveries, evidence, category spending and likely-needed products.
• Evidence screens that expose source receipts, comparison logic and claim limitations.
• Purchase-safe nutrition language: no diagnosis, calorie claim or assumed consumption.
• Local-first demo mode that works without credentials or a deployed backend.
• A production backend template using the OpenAI Responses API with structured receipt extraction. Its configured default is gpt-5.6; API credentials never enter the APK.
• Shared Kotlin Multiplatform domain contracts for future iOS work.

How we used Codex and GPT‑5.6

The product owner set the problem, tone, user safeguards and evidence standard: ANT must be a gentle mirror, not a scorecard. Codex accelerated implementation and review across the Android architecture, receipt-review flow, deterministic evidence model, tests, accessibility checks, release audit and submission materials.

GPT‑5.6 is configured in the production receipt-extraction backend as the structured multimodal model. The submitted APK deliberately runs a transparent synthetic local flow so judges can test it without credentials or network access. Before submitting, replace the placeholders below only with the real Codex thread evidence and /feedback Session ID.

Submission fields to complete before Devpost

Field	Value
Public repository	github.com/sqrl13/ant-purchase-intelligence
Public Android test build	Download ANT Build Week Demo v1.0.0
Public YouTube demo (< 3 minutes, English audio)	REPLACE_WITH_VIDEO_URL
Codex /feedback Session ID	019f61f9-caae-7722-a7b2-ca5621e24106
Codex model shown by the qualifying thread	REPLACE_WITH_VERIFIED_GPT_5_6_EVIDENCE

Do not state that a particular Codex model was used unless the qualifying thread proves it. The backend configuration alone does not prove the model used by a Codex session.

Build from source

Requirements: JDK 17, Android SDK 36 and Android Studio compatible with AGP 9.2.x.

.\gradlew.bat :app:testDebugUnitTest
.\gradlew.bat :app:lintDebug
.\gradlew.bat :app:assembleDebug

To package the test build and SHA-256 manifest:

powershell -ExecutionPolicy Bypass -File .\scripts\package-hackathon-apk.ps1

To run the full repository checks:

powershell -ExecutionPolicy Bypass -File .\scripts\verify-hackathon.ps1

Repository map

Path	Purpose
app/	Android app: Compose UI, Room data layer, receipt flow and tests.
shared/	Kotlin Multiplatform domain contracts and rules.
composeApp/, iosApp/	KMP/iOS foundation for the long-term product.
backend/	Optional production backend template; never needed for the offline demo.
docs/HACKATHON_SUBMISSION.md	Copy-ready English Devpost description and testing instructions.
docs/HACKATHON_VIDEO_RUNBOOK.md	Exact capture route and video safety checklist.
docs/HACKATHON_COMPLIANCE.md	Completed safeguards and entrant-owned external steps.
docs/HACKATHON_BUILD_LOG.md	Meaningful Build Week extension record.
SUBMISSION_MANIFEST.md	Exact repository contents and Devpost items submitted outside the repository.
docs/PUBLISHING_HANDOFF.md	Exact GitHub Release and YouTube upload handoff.

Safety and privacy

Receipt images remain local in the demo. Insights use confirmed purchases only. ANT never treats an extraction as truth, never embeds an AI key in the mobile app, and never turns a purchase into a medical or moral claim.

Licence

ANT is released under the MIT License. It permits use, modification and distribution of this source code while retaining the copyright and licence notice.