[ Tryout Project · polaris-image-2-photo-intelligence ]

Polaris Image 2 Photo Intelligence

Build a lightweight property photo pipeline that turns a full real-estate photo tour into schema-valid JSON with exact photo ID coverage.

Large1 to 2 focused weeks$500 bounty on approval

Submit by email Open project repo

Problem

Real-estate photo tours are messy: exterior shots, interiors, smartphone photos, professional photos, partial views, blurry images, and duplicate-ish sequences all arrive together.

A useful system must classify every photo, group photos into displayable spaces, choose hero images, and emit structured output without dropping, duplicating, or inventing photo IDs.

This public tryout is a sanitized version of a real Covent problem. The goal is to show that you can build a runnable pipeline with validation, evaluation, and honest failure analysis.

Scope

Accept a JSON input with `property_id` and `photos[]`.
Emit top-level `photos`, `spaces`, `property`, and `warnings` JSON.
Classify per-photo fields including room role, condition, image quality, and capture source.
Group photos into spaces and choose listing/per-space heroes.
Include deterministic validation for required fields, enum values, and photo ID references.
Include at least one sample run and an evaluation/writeup explaining quality, latency, and cost tradeoffs.

Acceptance criteria

Meet every line below and the submission is approved. The $500 bounty is paid, regardless of any internship decision.

Runs from a clean checkout with documented setup and run commands.
Passes the public validator in the tryout repo for the included sample input/output contract.
Preserves every input photo ID exactly once in `photos` and display-space coverage, with zero missing, duplicate, or hallucinated IDs.
Includes a non-trivial full-tour sample output and an evaluation report with metrics plus failure analysis.
Clearly explains whether the system is cheap-path only, VLM-only, or hybrid routed, and why.
Does not include private Covent data, secrets, raw customer images, or proprietary prompt artifacts.

Writeup

The artifact is half. The writeup is the other half. Show us how you think.

Explain the architecture and what each stage owns.
Describe how you validated schema correctness and photo ID coverage.
Describe your evaluation set, metrics, and biggest observed failure modes.
Estimate cost and latency, especially if you use hosted VLM calls.
List what you would improve with another week.

Assets

Project brief and public validator
Public-safe project brief, sample input/output, schema contract, and lightweight validator.

How to submit

Code submissions: fork the tryouts mono-repo, build inside projects/polaris-image-2-photo-intelligence/submissions/<your-handle>/, and open a PR with your code plus a WRITEUP.md.
Non-code submissions (design, ops, data): email talent@getcovent.com with subject Internship Tryout: polaris-image-2-photo-intelligence: <Your Name>. Attach a link to the artifact (repo, live URL, Figma, Notion, video, deck) plus your writeup.
Review: founders read every submission. We respond within 7 days.

Bounty rules

Every project has clear acceptance criteria. Meet the bar, get paid.
Uncapped: multiple people can solve the same project and each is paid.
The bounty is independent of any internship offer.
We respond within 7 days.
Payment via Wise (international) or ACH (US) after a W-9 or W-8BEN.
One bounty per person per project.

Covent is an equal opportunity employer. We hire on merit and fit, not background.

Problem

Real-estate photo tours are messy: exterior shots, interiors, smartphone photos, professional photos, partial views, blurry images, and duplicate-ish sequences all arrive together.

A useful system must classify every photo, group photos into displayable spaces, choose hero images, and emit structured output without dropping, duplicating, or inventing photo IDs.

This public tryout is a sanitized version of a real Covent problem. The goal is to show that you can build a runnable pipeline with validation, evaluation, and honest failure analysis.

Scope

Accept a JSON input with `property_id` and `photos[]`.

Emit top-level `photos`, `spaces`, `property`, and `warnings` JSON.

Classify per-photo fields including room role, condition, image quality, and capture source.

Group photos into spaces and choose listing/per-space heroes.

Include deterministic validation for required fields, enum values, and photo ID references.

Include at least one sample run and an evaluation/writeup explaining quality, latency, and cost tradeoffs.

Acceptance criteria

Meet every line below and the submission is approved. The $500 bounty is paid, regardless of any internship decision.

Runs from a clean checkout with documented setup and run commands.

Passes the public validator in the tryout repo for the included sample input/output contract.

Preserves every input photo ID exactly once in `photos` and display-space coverage, with zero missing, duplicate, or hallucinated IDs.

Includes a non-trivial full-tour sample output and an evaluation report with metrics plus failure analysis.

Clearly explains whether the system is cheap-path only, VLM-only, or hybrid routed, and why.

Does not include private Covent data, secrets, raw customer images, or proprietary prompt artifacts.

Writeup

The artifact is half. The writeup is the other half. Show us how you think.

Explain the architecture and what each stage owns.

Describe how you validated schema correctness and photo ID coverage.

Describe your evaluation set, metrics, and biggest observed failure modes.

Estimate cost and latency, especially if you use hosted VLM calls.

List what you would improve with another week.

How to submit

Code submissions: fork the tryouts mono-repo, build inside projects/polaris-image-2-photo-intelligence/submissions/<your-handle>/, and open a PR with your code plus a WRITEUP.md.

Non-code submissions (design, ops, data): email talent@getcovent.com with subject Internship Tryout: polaris-image-2-photo-intelligence: <Your Name>. Attach a link to the artifact (repo, live URL, Figma, Notion, video, deck) plus your writeup.

Review: founders read every submission. We respond within 7 days.

Bounty rules

Every project has clear acceptance criteria. Meet the bar, get paid.

Uncapped: multiple people can solve the same project and each is paid.

The bounty is independent of any internship offer.

We respond within 7 days.

Payment via Wise (international) or ACH (US) after a W-9 or W-8BEN.

One bounty per person per project.