What Is a Research Hackathon? 5 AI Projects From Paris!

One difficult AI problem can produce several credible solutions. Testing them sequentially may take months. A research hackathon allows independent teams to investigate those directions at the same time, under the same deadline and against the same benchmark.

That is exactly what happened at the Iterate Paris Research Hackathon. Four teams tackled a cross-modal brain MRI retrieval challenge, each using a different combination of image registration, mutual information, modality-invariant descriptors, and robust preprocessing. A fifth team developed a generative model capable of turning a simple apartment outline into complete vector room polygons.

The event produced more than five technical demos. Competing methods became directly comparable, foundation-model baselines were challenged, a potential data leak was identified, and one training-free approach achieved perfect retrieval accuracy across all three MRI datasets.

Iterate designs and runs research hackathons for AI companies, from challenge definition and researcher sourcing to technical evaluation and final results. The five projects from Paris show how parallel experimentation can turn an unresolved research question into measurable evidence.

What Is a Research Hackathon?

A research hackathon is a time-boxed event where several teams investigate a defined technical question. Participants develop hypotheses, build experimental pipelines, test different methods, and report their results against shared evaluation criteria.

The objective differs from a conventional product hackathon. A polished interface or persuasive demo is not enough. Each team must explain what it tested, how the experiment was evaluated, and whether the evidence supports its original hypothesis.

Traditional research	Research hackathon
One team explores a limited number of directions	Several teams test competing approaches
Experiments often run sequentially	Experiments run in parallel
Evaluation may vary between projects	Teams use a shared benchmark
Research remains inside one organization	External researchers and engineers contribute
Findings emerge over a longer cycle	Initial findings appear within a fixed event

💡 A research hackathon does not replace peer review, extended testing, or production validation. It accelerates the exploratory stage by identifying promising methods, exposing weak assumptions, and eliminating approaches that fail under controlled conditions.

Negative results also create value. A team may demonstrate that a complex model underperforms a simpler baseline, detect a flaw in the dataset, or establish that a popular technique does not transfer to the selected problem. Each outcome narrows the search space for the research that follows.

Inside the Iterate Paris Research Hackathon

The Iterate Paris Research Hackathon placed measurable experimentation ahead of product polish. Five teams developed and evaluated original methods, with four working on the same cross-modal brain MRI retrieval problem.

This shared challenge created an unusual research environment. Instead of presenting unrelated projects, the teams produced competing solutions that could be examined side by side. Each group worked with the same underlying task while making independent decisions about image representation, registration, similarity measurement, and preprocessing.

The submissions explored five distinct directions:

✅ VoxelBridge: GPU registration combined with mutual information;

✅ CrossMIND: Modality-invariant MIND descriptors and registration;

✅ Alignment-First Matching: Affine alignment before similarity scoring;

✅ Yantra MRI Research: SSC descriptors, content registration, and resection-robust trimming;

✅ outlineFlow: Vector floor-plan generation with rectified flow matching.

The first four projects tested whether cross-modal MRI retrieval required a larger neural architecture or a better representation of anatomical alignment. outlineFlow addressed a separate generative problem, showing how the same event format could also support from-scratch model development and iterative performance improvements.

By giving independent teams a defined problem and measurable evaluation criteria, Iterate turned the event into a parallel research environment. The result was not one proposed solution, but several technical approaches, documented experiments, and directly comparable outcomes.

The Shared Challenge: Matching Patients Across MRI Modalities

The central challenge asked teams to retrieve scans belonging to the same patient across different MRI modalities. This is difficult because a brain can look significantly different depending on how the image was captured.

T1, T2, and contrast-enhanced T1 scans emphasize different anatomical properties. Fluid, tissue, lesions, and resection areas may appear with different intensities. As a result, two scans of the same patient can look less similar at the pixel level than scans from two different patients.

Teams also had to account for geometric variation. The head may be positioned differently, the image may use another scale or orientation, and surgery may have altered part of the anatomy. A retrieval system therefore needs to distinguish meaningful anatomical identity from changes caused by modality or alignment.

The task followed a straightforward retrieval process:

1. Take one MRI scan as the query.
2. Compare it with a set of candidate scans from another modality.
3. Rank the candidates by similarity.
4. Measure how highly the correct patient appears.

Performance could then be evaluated through metrics such as:

• Retrieval accuracy: The percentage of queries for which the correct patient is successfully identified.
• Mean Reciprocal Rank (MRR): A ranking metric that gives a higher score when the correct match appears near the top of the results.
• Macro MRR: The average MRR calculated across datasets or evaluation groups so that each group contributes equally.

The obvious approach was to use a large pretrained model to extract semantic features from every scan. However, several teams reached a different conclusion: cross-modal retrieval depended less on abstract semantic understanding than on correctly aligning the underlying anatomy. This reframing shaped the strongest submissions.

Four Teams, Four Approaches to Brain MRI Retrieval

The four teams worked on the same cross-modal retrieval problem but selected different representations and alignment methods. All four developed training-free pipelines, making their results especially useful for direct comparison.

Project	Core approach	Training required	Reported result
VoxelBridge	GPU registration and mutual information	No	1.000 retrieval accuracy
CrossMIND	MIND descriptors and GPU registration	No	0.703 macro MRR
Alignment-First Matching	Mutual information and affine registration	No	Approximately 0.96 MRR
Yantra MRI Research	SSC, content registration, and robust trimming	No	Not specified

#1 VoxelBridge: Perfect Retrieval Without Deep Learning

VoxelBridge achieved first place with a training-free pipeline designed to match the same patient across brain MRI modalities.

The method followed two main stages:

1. Anatomical alignment: GPU-accelerated registration aligned the scans before comparison.
2. Similarity scoring: Mutual information measured the correspondence between each aligned pair.

The system achieved 1.000 retrieval accuracy across all three datasets, outperforming the other reported approaches without using deep learning.

This result supports the project’s central hypothesis: cross-modal differences did not necessarily require a learned representation when the underlying anatomy could first be aligned effectively.

Built by Team TakeTheMoneyAndRun—Amar Ribic, Alankrit Sharma, Lukas Sokcevic, Philip Thome, and Leon Ritz—VoxelBridge also provided the strongest example of a research hackathon testing a simpler alternative to model-heavy pipelines. The winning method focused on the structure of the problem rather than adding architectural complexity.

#2 CrossMIND: Modality-Invariant Descriptors for MRI Retrieval

CrossMIND also used a training-free pipeline, but replaced mutual information with modality-independent neighborhood descriptors, or MIND descriptors.

These descriptors represent local anatomical structures while reducing the influence of contrast differences between ceT1 and T2 scans. Team Harissa combined them with GPU registration to align the images before comparing their representations.

The pipeline followed three main stages:

1. Register the scans to improve anatomical alignment.
2. Extract MIND descriptors from the aligned images.
3. Rank the candidates according to descriptor similarity.

CrossMIND achieved a macro MRR of 0.703, representing a 55% improvement over the baseline.

Built by Fares Elkholy, Amine Cheikhrouhou, Mohamed Kechiche, Firas Mosbeh, and Mohamed Yacine Dhouafli, the project tested whether modality-invariant local descriptors could support patient matching without model training. Its results provided a second piece of evidence that preprocessing and anatomical representation could outperform a more model-dependent retrieval pipeline.

#3 Alignment-First Matching: Treating Retrieval as a Geometry Problem

Team Laurence reframed cross-modal brain MRI retrieval as an alignment problem rather than a deep-learning problem.

The team found that mutual information could match corresponding T1 and T2 scans when the images were already aligned. When geometric differences disrupted that correspondence, affine registration restored the alignment before similarity scoring.

The approach therefore used two complementary components:

• Affine registration to correct geometric misalignment;
• Mutual information to compare the aligned scans across modalities.

This training-free pipeline increased the reported MRR from approximately 0.20 with foundation-model baselines to approximately 0.96.

Built by Kellian VK, Katerina Michenina, Laurence Liang, Alvaro Serero, and Tina Ranich, the project challenged the assumption that stronger learned representations were required. Under the conditions tested, correcting the geometry before measuring similarity produced a substantially better retrieval result.

#4 Yantra MRI Research: Building a Robust, Leak-Free Pipeline

Team Yantra developed a deterministic, training-free system for cross-modal 3D brain MRI patient re-identification.

The pipeline combined three components:

• Modality-invariant SSC descriptors to represent anatomical content across different MRI contrasts;
• Content registration to align corresponding structures;
• Resection-robust trimming to reduce the influence of surgically altered regions.

The team tested and mapped more than 20 alternative techniques that failed to outperform its final approach. This negative evidence documented which directions did not improve retrieval instead of presenting only the successful pipeline.

Yantra also identified a planted geometric leak in the data. Exploiting it could have produced a stronger score without solving the intended patient-matching problem. The team measured the leak, reported it, and refused to use it as part of the final method.

Built by Gowshigan Selladurai, Wilfred Doré, Anshu Raj, Abdul Rafay Mohammed, and Frederik Tress, Yantra MRI Research demonstrates that benchmark integrity matters as much as benchmark performance. A credible research result must reflect the intended task rather than an accidental shortcut hidden in the data.

What Four Parallel Experiments Revealed

The four teams started from the same cross-modal MRI retrieval problem and independently converged on one broad conclusion: better anatomical representation and alignment mattered more than adding a larger learned model.

Several patterns emerged across the submissions:

✅ Every final approach was training-free. None of the four teams needed to fine-tune a deep-learning model for the retrieval task.

✅ Registration appeared repeatedly. GPU, affine, or content registration formed part of every proposed pipeline.

✅ Modality differences required deliberate handling. Teams used mutual information, MIND descriptors, or SSC descriptors rather than relying on direct image similarity.

✅ Simpler methods competed strongly with learned baselines. VoxelBridge reported perfect retrieval accuracy, while Alignment-First Matching increased MRR from approximately 0.20 to approximately 0.96 compared with its foundation-model baselines.

✅ Data quality influenced the validity of the results. Yantra’s detection of a geometric leak showed how a benchmark can reward a shortcut instead of the intended solution.

These findings do not establish that training-free methods will outperform deep learning across every medical imaging task. They reflect the datasets, baselines, and evaluation conditions used during the hackathon.

Their convergence remains significant. Instead of receiving one recommendation from one research team, the event produced four independent experiments pointing toward alignment-first methods. That is the central advantage of parallel research: repeated findings become visible within the same evaluation cycle.

outlineFlow: Generating Vector Floor Plans From an Outline!

While the other teams focused on brain MRI retrieval, outlineFlow investigated a separate generative research problem: creating a complete apartment layout from its exterior outline.

Team Lihua built a rectified flow-matching set-Transformer from scratch. Given only the apartment boundary, the model generates a set of typed room polygons, including both their shape and position.

The system produces vector geometry rather than a pixel-based floor-plan image. This makes each generated room a structured polygon that can be edited, measured, or passed into another design workflow.

During development, the team identified an L2 mean-collapse problem that limited the diversity and quality of the generated layouts. It then introduced two changes:

• Grid-align to improve the spatial structure of the output;
• Churn sampling to improve generation during the sampling process.

Following these adjustments, outlineFlow reported:

• FID improvement from 167 to 135
• Coverage increase of 66%

Built by Bouyahia, Hohamed Aymen, Jiahua Tang, Gaia Mezaib, and Kaiying Wu, outlineFlow shows another function of a research hackathon: teams can diagnose a measurable model failure, modify the sampling pipeline, and validate the effect within the same research cycle.

Run a Research Hackathon With Iterate!

A research backlog can leave several credible technical directions waiting for internal resources. Iterate turns a defined research problem into a structured event where selected AI researchers and engineers test those directions in parallel.

Iterate manages the hackathon end to end:

👉 Research challenge scoping

👉 Builder and researcher sourcing

👉 Application review and team formation

👉 Event operations

👉 Technical project evaluation

👉 Final demos and result presentation

👉 Identification of high-performing technical talent

This model gives AI companies access to more than isolated project submissions. Every team produces evidence of how it frames an uncertain problem, selects an approach, responds to failed experiments, and delivers under a fixed deadline.

The resulting projects can reveal promising research directions, expose weaknesses in an existing benchmark, and create a pipeline of engineers already tested on relevant technical work.

Run a research hackathon with Iterate to turn an unresolved AI problem into parallel experiments, measurable results, and direct access to proven builders.

Conclusion

The Paris Research Hackathon turned one MRI retrieval challenge into four independent experiments, while outlineFlow tested an entirely different generative approach. The results show how parallel research can compare methods, expose benchmark flaws, and identify promising directions faster.

Iterate runs research hackathons that connect difficult AI problems with researchers and engineers capable of turning hypotheses into measurable results.

Frequently Asked Questions

What is the difference between a research hackathon and a regular hackathon?

A regular hackathon usually prioritizes working products and compelling demos. A research hackathon focuses on technical hypotheses, experimental methods, shared benchmarks, and measurable findings.

What can a research hackathon produce?

Outputs may include experimental prototypes, benchmark results, reusable code, unsuccessful methods, dataset findings, and new research directions. The event can also identify engineers who perform well on open technical problems.

Can a research hackathon replace a conventional research project?

No. It accelerates exploration and helps identify promising approaches, but the results still require extended testing, validation, and potentially peer review before broader scientific or production use.

How does Iterate organize research hackathons?

Iterate manages challenge scoping, researcher sourcing, applications, team formation, event operations, project evaluation, and final demos. AI companies receive several tested approaches and direct access to builders who have already worked on the selected problem.