Building a 29-Country Home Cook Network for AI Food Image Data

Building a 29-Country Home Cook Network for AI Food Image Data

A leading AI research and data company needed thousands of authentic home-cooked meals collected across 29 countries for AI training. Each meal had to be documented through a strict process that included the recipe, ingredient photographs on a kitchen scale, a plated image and video, and weighed servings.

This was not simply an image collection project. The operation had to recruit the right contributors, verify that submissions were genuinely created in the assigned market, manage meal reservations, maintain consistency across thousands of submissions, and process a large quality-review workload before the data reached the client.

Over roughly six weeks of production, the project grew to more than 6,300 registered contributors, 1,500+ active submitters, 18 partner agencies, and a 34-member QC team.

The result was 11,400+ meals submitted, more than 7,200 meals collected after internal QC, and 6,330 meals accepted by the client, with a 91% acceptance rate.

The Requirement

The client needed authentic home-cooked meal data from contributors physically based in 29 countries.

Every submission had to meet a structured collection protocol. Contributors were expected to prepare the meal themselves and document the process through recipes, ingredient photographs, weighing evidence, plated images, and video.

The requirement therefore combined several operational needs at once: high-volume contributor recruitment, country-level authenticity, duplicate prevention, structured submission management, quality assurance, and ongoing client delivery.

The project also had to move quickly. Each country had its own target, and those targets could change as production progressed.

Why the Project Was Difficult

The first challenge was finding the right contributors.

The ideal participant was often not a professional freelancer or content creator. They were everyday home cooks with access to a smartphone and kitchen scale who were willing to follow a detailed collection process.

That meant recruitment had to extend beyond traditional freelance networks and adapt to very different markets.

The second challenge was authenticity. Every meal had to be prepared in the contributor's assigned country. VPN use was not permitted, and repeat dishes had to be prevented.

At small scale, these controls could potentially be handled manually. At global scale, that approach would quickly become unreliable and expensive.

The third challenge was volume.

At peak production, as many as 718 meals arrived in a single day. Each meal included approximately 16 files requiring review. That meant thousands of individual assets could enter the QC queue in a single day.

The operation therefore had to increase throughput without allowing quality standards to weaken.

The final challenge was changing market demand. Country targets shifted during production, which meant individual markets sometimes had to be opened, paused, or closed quickly.

The project needed to remain responsive without disrupting contributors or allowing unnecessary collection to continue after a target had been reached.

The Approach

The project was managed as one connected production pipeline rather than a collection of separate sourcing, submission, and review activities.

One Platform for the Full Workflow

A dedicated portal managed contributor sign-up, dish reservations, guided uploads, internal QC, client delivery, and payouts.

Country-level targets were also tracked through the same workflow.

This created a single operational view of the project and made it easier to understand how contributors and submissions were moving through the system.

It also allowed the team to react more quickly as individual markets approached their targets.

Multi-Channel Contributor Recruitment

A project spanning 29 countries could not depend on one sourcing channel.

Contributors were recruited through freelance platforms, local outreach, and partner networks. 18 partner agencies were brought into the project to strengthen coverage in harder-to-reach markets.

Where appropriate, nearby countries with shared regional cuisines were also used to widen the available contributor pool.

This gave the project flexibility to use different sourcing models depending on the difficulty of each market.

Verifying Contributors and Protecting Authenticity

Location verification was built into the process.

Device-location and IP checks were used to help confirm that contributors were operating from the correct country, while VPN use was rejected.

Duplicate prevention was also automated.

More than 13,900 dish reservations were screened using an AI-based duplicate check to reduce the risk of repeated dishes entering the collection.

This helped protect dataset variety before contributors had invested time preparing and submitting the meal.

Two-Level Quality Control

Quality assurance operated as a production line of its own.

A 34-member review team worked from a shared queue, supported by a second-level quality check.

Where submissions could be corrected, contributors received clear reasons for rework along with a defined window to make the necessary changes.

This created a more productive feedback loop than simply rejecting every imperfect submission.

The scale of this operation became substantial. More than 27,800 QC reviews were completed across the project, covering over 184,000 photos and videos.

On the busiest review day, the QC team completed 2,038 reviews.

Connecting Internal QC With Client Feedback

Internal review was not the final step.

Client verdicts were synced back into the workflow every 10 minutes, giving the operations team an updated view of acceptance outcomes.

This allowed decisions to be made using actual project status.

Markets approaching completion could be slowed down or closed. Quality issues could be identified more quickly. Accepted meals could move into payout without unnecessary delay.

The feedback loop connected production directly with client acceptance.

The Result

In approximately six weeks of production, the project generated more than 11,400 submitted meals.

Over 7,200 meals were collected after internal QC, and 6,330 meals were accepted by the client.

The resulting client acceptance rate was 91%.

The first major checkpoint of 6,815 meals was reached in under four weeks, with 26 of 29 countries already past their individual checkpoint at that stage.

The supporting QC operation completed more than 27,800 reviews across 184,000+ photos and videos.

Headline Metrics

• 29 countries covered

• 6,300+ contributors registered

• 1,500+ active submitters

• 18 partner agencies

• 11,400+ meals submitted

• 7,200+ meals collected after QC

• 6,330 client-accepted meals

• 91% client acceptance rate

• 34 QC reviewers

• 27,800+ QC reviews completed

• 184,000+ photos and videos reviewed

• 718 meals received on the peak production day

• 2,038 QC reviews completed on the peak review day

Why the Model Worked

The strongest part of the project was that contributor sourcing, submission management, quality review, and client feedback were treated as parts of the same system.

Multi-channel recruitment created enough supply across 29 countries.

Location checks and duplicate screening helped protect authenticity.

Two-level quality control allowed review capacity to scale alongside submission volume.

Live country tracking made it possible to respond to changing targets instead of continuing collection blindly.

And client feedback integration meant that operational decisions could be based on actual acceptance outcomes rather than delayed reporting.

The project scaled because each stage of the workflow was designed to support the next.

Key Takeaway

Large-scale AI data collection is not simply about recruiting more contributors.

The real challenge is converting contributor activity into usable, verified, client-accepted data.

For this project, that required a system capable of coordinating recruitment across 29 countries, contributor verification, reservation management, tens of thousands of QC decisions, live market targets, client feedback, and payouts at the same time.

The result was a global home-cook network built in weeks, supported by an operational model designed to balance volume, authenticity, quality, and speed.