Each session runs about two weeks and can be joined on its own. In those two weeks, students complete a project with a Project Lead, visit research institutions, and meet researchers and founders working in Boston. Here's what this summer looked like.

01 · Who joined

This summer's students came from South China University of Technology, Peking University, Xi'an Jiaotong University, the University of Tokyo, the National University of Singapore, Universidad Panamericana, the University of Manchester, and New England Conservatory, among others.

They study computer science, data science, engineering, finance, biology, geography — and music. Some had just finished their first year of undergrad; others were master's students or recent graduates.

02 · Four PBLs, and what students built

Students worked in small teams on Project-Based Learning (PBL). Each project is set in a real industry scenario, led by a Project Lead, and ends with a public presentation.

Each PBL had more teams and presentations than the ones shown here. The four below are a selection — a part of the work from this summer, not a ranking of the best projects.

AI and Computer Vision in BioTech – Novo Nordisk Project

Brain tumor segmentation and volume measurement

The team combined a CNN with a vision transformer to outline brain tumors in MRI scans, then wrapped the model in a web app that accepts MRI uploads.

• Mean Dice score 0.8481 on 73 patients held out from training (1.0 = a perfect match with the clinician's annotation)

• Mean volume error 12.3 cm³

AI and Robotics for Mobile Robot Manipulation – Boston Dynamics Project

3D path planning for a mobile robot

The team compared two path-planning algorithms in simulation.

• In an S-shaped corridor, Informed RRT* succeeded in 4 of 5 runs; RRT* succeeded in 2

• Conclusion: the right algorithm depends on whether the task prioritizes speed, stability, or path quality

AI in Operations Strategy – Amazon Project

Once the investor's rules are set, does the risk model still matter?

Using 2016–2025 data, the team compared five risk models under five sets of investor rules.

• Sharpe ratios across models landed within 0.04 of each other

• None clearly beat an equal-weight portfolio — what mostly determined risk was the investor rules themselves

AI and Data Visualization for Social Impact – Tableau Project

Which users does a voice assistant fail most often?

The team analyzed recognition failures across accents, native languages, and usage habits. Each student took one question, and the pieces came together in an interactive dashboard now published on Tableau Public.

Final presentations

At the end of the two weeks, every team presented working code and models, an interactive dashboard or web tool, and a project poster — plus a clear account of the problem, the work, and the results.

In the audience were the Project Leads and researchers and founders from research institutions, companies, and startups. They asked about methods and results on the spot, and students answered live.

03 · Meeting researchers and founders

Beyond the projects, students attended two community events.

AI+X Meetup at the MIT Museum

Guests that evening (titles as of the event):

• AI Research Scientist, Lila Sciences — research at MIT using generative AI to propose candidate materials and the experimental procedures to synthesize them. His point: scientists still have to stay at the center of decisions, judgment, and validation.

• Senior Director of AI & Machine Learning, Moderna — in high-stakes settings like drug development, AI systems need clear roles, reviewed outputs, and traceable actions. The stronger the AI, the more reliability and accountability matter.

• Founder & CEO, Celvion Therapeutics — a platform that starts from real clinical outcomes to find more promising cancer combination therapies.

• Researcher, MIT Mechanical Engineering & Computational Science — physics-informed AI to test vehicle crash designs more efficiently and catch likely failure modes early.

• Co-founder, Openbase — a live demo of one person directing multiple AI agents by voice.

• Director, Venture Café Cambridge — as AI gets stronger, real in-person connection gets scarcer.

"Show up in life being curious." — Director, Venture Café Cambridge

Four students then took the stage:

• Karan R., National University of Singapore — President of the NUS Business Analytics Consulting Team, on how his student organization connects to the wider AI+X network.

• Nam P., University of Tokyo — from the University of Tokyo to Bosch Japan, Audi, and CERN: finding direction across research and industry.

• Daniela V., Universidad Panamericana — data science, computer vision, biotech, robotics, and social impact, from a Women in Tech perspective.

• Monica C., New England Conservatory — a cellist who has just started learning to code, on why AI isn't only for CS majors.

AI+X Founders × Researchers at Microsoft New England

Co-hosted by BlendED and TEDxHarvard Square. Attendees applied by describing the project or research they're working on and were reviewed before being admitted — about 230 people applied.

A few founders and researchers gave short introductions to their work, and most of the evening was left for conversation. Past attendees of this series have come from MIT Media Lab, Harvard Innovation Labs, The Engine, Autodesk, LabCentral, Greentown Labs, and MassChallenge. OCE students attended as participants, asking questions and talking with founders and researchers.

04 · Site visits and fireside chats

Broad Institute

Students visited the Broad Institute of MIT and Harvard, the genomics and biomedical research institute in Kendall Square.

Fireside chats

Small-group Q&A: the host opens with a few questions, then students ask freely. Most took place in the Quick Room on the MIT campus.

• Agentic AI Solutions Architect — how enterprise AI projects actually move forward, and what skills the roles require.

• Researcher, MIT Media Lab Camera Culture Group — why he chose research, and how it differs from industry work. Most questions came from students considering a PhD.

• Researcher, MIT Mechanical Engineering & Computational Science — how AI is entering the engineering design process.

• Assistant Professor of Computer Science, Boston College — data visualization and human-computer interaction, and how cultural and educational background shapes the career choices of international scholars.

• Postdoctoral Fellow, Eric and Wendy Schmidt Center at the Broad Institute — combining physical models with AI for biomedical problems like blood glucose dynamics.

AI+X Founder Dinner

BlendED and Rho, a fintech platform for startups, co-hosted a dinner in Kendall Square with 14 early-stage founders working in medical AI, biotech, fintech, and aerospace hardware. Among them:

• Haver AI — AI clinical decision support and governance for healthcare organizations

• Deliverome Bio — mapping cell-surface proteins for targeted drug delivery

• EON Health — a longevity AI platform built on wearable and health data

• TaxScout.ai — AI that processes tax documents for accounting firms

• Ballista Labs — an AI agent that designs aircraft piping inside engineering CAD software

• Synquery — AI that matches companies with industry experts

• Openbase — voice control for multiple AI agents

BlendED shared three ways founders can work with the community: recruiting global talent, posting collaborative projects, and using the community to reach global markets.

05 · Boston and Cambridge after class

• Charles River runs — Saturday mornings, August 8 and 22

• Harvard Square — some students stayed in Lesley University housing nearby and spent most evenings around the area

• Harvard Business School campus

• The Freedom Trail — weekend walks through colonial and Revolutionary-era sites

• Fenway Park — a Red Sox game

06 · Student voices

"It's like real research: a one-to-two-hour meeting each week to go over the week's key progress, raise problems, solve them, figure out what doesn't work, and come back with better questions."

— Yeshwanth P., University of Manchester · Computer Science

Completed online projects in robot path planning and 3D scene representation, and shared his experience at an info session in China.

"Before I came, I knew almost nothing about AI and worried I'd be far behind everyone. Then I realized the program is there to teach you AI — it doesn't expect you to know it from day one."

— Edwin W., Penn State · Finance

Joined two on-campus PBLs, in data visualization and robotics.

"Even if you're not confident in your English or in learning a new field, I really recommend coming to Boston, because you get to work with students from very different backgrounds."

— Kenta U., Tokyo Metropolitan University · Master's in Computer Science (Cybersecurity)

Six months online, then two OCE sessions on campus.

"An on-campus session is only 14 days, but you can use those 14 days to work closely with your Project Lead."

— Student, Xi'an Jiaotong University · Mechanical Engineering

Joined four on-campus PBLs and shared his experience at an info session in China.

07 · 2027 OCE dates

Each session can be joined on its own.

→ Apply for the 2027 OCE

To evaluate reconstruction quality, calibration and reconstruction diagnostic tools are used to examine reprojection errors, checkerboard coverage, stereo-pose consistency, rectification quality, valid-depth coverage, and disparity and depth distributions. The project also investigates multi-view reconstruction by capturing objects from different vertical viewing angles to recover additional surfaces that may be occluded in individual views. Using Open3D, individual point clouds can be downsampled, filtered for outliers, and aligned through Iterative Closest Point (ICP) registration before being merged into a combined point cloud. Experimental testing highlights the importance of appropriate industrial-camera configuration, stable and consistent image acquisition, adequate lighting, sufficient surface texture, reliable calibration, and substantial overlap between viewpoints. Overall, the project demonstrates an end-to-end stereo reconstruction workflow while examining the practical challenges of obtaining coherent and sufficiently complete 3D representations from real-world image data.

The system compares a reference dance clip with a practice clip through a four-step workflow: clip selection, motion alignment, difference identification, and targeted review. It extracts human body keypoints, normalizes poses around the hips and torso, and uses Dynamic Time Warping (DTW) to align movement phases performed at different speeds. Joint-angle comparisons provide a similarity summary, while additional diagnostics report body-region differences, angular velocity, limb direction, observation coverage, and suggested review windows.

The project evolved through three main versions. V1.0 established the baseline comparison pipeline. V1.2 improved comparison stability through sequence-level torso-scale normalization, angle smoothing, and score mapping, while providing body-part and timing reports. V1.3 retained the V1.2 score and extended the system with more interpretable diagnostic information for review, including velocity, limb-direction, left/right-region, and data-quality indicators.

We also prepared a local interactive prototype with precomputed dance examples. Users can compare aligned videos, view skeleton overlays, inspect regional differences, and jump to suggested review windows. The system is intended to support human feedback rather than replace a dance teacher or provide an automatically validated performance grade.

The project’s current impact is a clearer bridge between technical motion metrics and practical dance review. It helps users ask not only “How similar are the clips?” but also “Where should I look?”, “When should I review it?”, and “Can this observation be trusted?” Future work includes natural-language feedback, on-video annotations, expert-labeled evaluation, and user studies with dancers and teachers.

The system compares a reference dance clip with a practice clip through a four-step workflow: clip selection, motion alignment, difference identification, and targeted review. It extracts human body keypoints, normalizes poses around the hips and torso, and uses Dynamic Time Warping (DTW) to align movement phases performed at different speeds. Joint-angle comparisons provide a similarity summary, while additional diagnostics report body-region differences, angular velocity, limb direction, observation coverage, and suggested review windows.

The project evolved through three main versions. V1.0 established the baseline comparison pipeline. V1.2 improved comparison stability through sequence-level torso-scale normalization, angle smoothing, and score mapping, while providing body-part and timing reports. V1.3 retained the V1.2 score and extended the system with more interpretable diagnostic information for review, including velocity, limb-direction, left/right-region, and data-quality indicators.

We also prepared a local interactive prototype with precomputed dance examples. Users can compare aligned videos, view skeleton overlays, inspect regional differences, and jump to suggested review windows. The system is intended to support human feedback rather than replace a dance teacher or provide an automatically validated performance grade.

The project’s current impact is a clearer bridge between technical motion metrics and practical dance review. It helps users ask not only “How similar are the clips?” but also “Where should I look?”, “When should I review it?”, and “Can this observation be trusted?” Future work includes natural-language feedback, on-video annotations, expert-labeled evaluation, and user studies with dancers and teachers.

To evaluate reconstruction quality, calibration and reconstruction diagnostic tools are used to examine reprojection errors, checkerboard coverage, stereo-pose consistency, rectification quality, valid-depth coverage, and disparity and depth distributions. The project also investigates multi-view reconstruction by capturing objects from different vertical viewing angles to recover additional surfaces that may be occluded in individual views. Using Open3D, individual point clouds can be downsampled, filtered for outliers, and aligned through Iterative Closest Point (ICP) registration before being merged into a combined point cloud. Experimental testing highlights the importance of appropriate industrial-camera configuration, stable and consistent image acquisition, adequate lighting, sufficient surface texture, reliable calibration, and substantial overlap between viewpoints. Overall, the project demonstrates an end-to-end stereo reconstruction workflow while examining the practical challenges of obtaining coherent and sufficiently complete 3D representations from real-world image data.

The system compares a reference dance clip with a practice clip through a four-step workflow: clip selection, motion alignment, difference identification, and targeted review. It extracts human body keypoints, normalizes poses around the hips and torso, and uses Dynamic Time Warping (DTW) to align movement phases performed at different speeds. Joint-angle comparisons provide a similarity summary, while additional diagnostics report body-region differences, angular velocity, limb direction, observation coverage, and suggested review windows.

The project evolved through three main versions. V1.0 established the baseline comparison pipeline. V1.2 improved comparison stability through sequence-level torso-scale normalization, angle smoothing, and score mapping, while providing body-part and timing reports. V1.3 retained the V1.2 score and extended the system with more interpretable diagnostic information for review, including velocity, limb-direction, left/right-region, and data-quality indicators.

We also prepared a local interactive prototype with precomputed dance examples. Users can compare aligned videos, view skeleton overlays, inspect regional differences, and jump to suggested review windows. The system is intended to support human feedback rather than replace a dance teacher or provide an automatically validated performance grade.

The project’s current impact is a clearer bridge between technical motion metrics and practical dance review. It helps users ask not only “How similar are the clips?” but also “Where should I look?”, “When should I review it?”, and “Can this observation be trusted?” Future work includes natural-language feedback, on-video annotations, expert-labeled evaluation, and user studies with dancers and teachers.

The system compares a reference dance clip with a practice clip through a four-step workflow: clip selection, motion alignment, difference identification, and targeted review. It extracts human body keypoints, normalizes poses around the hips and torso, and uses Dynamic Time Warping (DTW) to align movement phases performed at different speeds. Joint-angle comparisons provide a similarity summary, while additional diagnostics report body-region differences, angular velocity, limb direction, observation coverage, and suggested review windows.

The project evolved through three main versions. V1.0 established the baseline comparison pipeline. V1.2 improved comparison stability through sequence-level torso-scale normalization, angle smoothing, and score mapping, while providing body-part and timing reports. V1.3 retained the V1.2 score and extended the system with more interpretable diagnostic information for review, including velocity, limb-direction, left/right-region, and data-quality indicators.

We also prepared a local interactive prototype with precomputed dance examples. Users can compare aligned videos, view skeleton overlays, inspect regional differences, and jump to suggested review windows. The system is intended to support human feedback rather than replace a dance teacher or provide an automatically validated performance grade.

The project’s current impact is a clearer bridge between technical motion metrics and practical dance review. It helps users ask not only “How similar are the clips?” but also “Where should I look?”, “When should I review it?”, and “Can this observation be trusted?” Future work includes natural-language feedback, on-video annotations, expert-labeled evaluation, and user studies with dancers and teachers.

To evaluate reconstruction quality, calibration and reconstruction diagnostic tools are used to examine reprojection errors, checkerboard coverage, stereo-pose consistency, rectification quality, valid-depth coverage, and disparity and depth distributions. The project also investigates multi-view reconstruction by capturing objects from different vertical viewing angles to recover additional surfaces that may be occluded in individual views. Using Open3D, individual point clouds can be downsampled, filtered for outliers, and aligned through Iterative Closest Point (ICP) registration before being merged into a combined point cloud. Experimental testing highlights the importance of appropriate industrial-camera configuration, stable and consistent image acquisition, adequate lighting, sufficient surface texture, reliable calibration, and substantial overlap between viewpoints. Overall, the project demonstrates an end-to-end stereo reconstruction workflow while examining the practical challenges of obtaining coherent and sufficiently complete 3D representations from real-world image data.

To evaluate reconstruction quality, calibration and reconstruction diagnostic tools are used to examine reprojection errors, checkerboard coverage, stereo-pose consistency, rectification quality, valid-depth coverage, and disparity and depth distributions. The project also investigates multi-view reconstruction by capturing objects from different vertical viewing angles to recover additional surfaces that may be occluded in individual views. Using Open3D, individual point clouds can be downsampled, filtered for outliers, and aligned through Iterative Closest Point (ICP) registration before being merged into a combined point cloud. Experimental testing highlights the importance of appropriate industrial-camera configuration, stable and consistent image acquisition, adequate lighting, sufficient surface texture, reliable calibration, and substantial overlap between viewpoints. Overall, the project demonstrates an end-to-end stereo reconstruction workflow while examining the practical challenges of obtaining coherent and sufficiently complete 3D representations from real-world image data.

To evaluate reconstruction quality, calibration and reconstruction diagnostic tools are used to examine reprojection errors, checkerboard coverage, stereo-pose consistency, rectification quality, valid-depth coverage, and disparity and depth distributions. The project also investigates multi-view reconstruction by capturing objects from different vertical viewing angles to recover additional surfaces that may be occluded in individual views. Using Open3D, individual point clouds can be downsampled, filtered for outliers, and aligned through Iterative Closest Point (ICP) registration before being merged into a combined point cloud. Experimental testing highlights the importance of appropriate industrial-camera configuration, stable and consistent image acquisition, adequate lighting, sufficient surface texture, reliable calibration, and substantial overlap between viewpoints. Overall, the project demonstrates an end-to-end stereo reconstruction workflow while examining the practical challenges of obtaining coherent and sufficiently complete 3D representations from real-world image data.