Flagging and Locating Anomalies in Retinal ScansOutcomeResearchAI + Health TechA team combined a Vision Transformer and a VQ-VAE to flag abnormal retinal scans and show where the problem is. The classifier reached 97.31% accuracy on 2,000 held-out scans./outcome-retinal-oct-anomaly-detectionAI and Computer Vision in BioTech – Novo Nordisk ProjectPresentation · PBL, May 2025180https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3faeae53dec3d65242c_thumb-v3-retinal-oct.jpg
Forecasting Induced Earthquakes with XGBoostOutcomeResearchAI+X ElectivesWorking alone, a student forecast earthquake rates in the Los Angeles Basin from well data. R² rose from 0.367 at 3 months to 0.842 at 5 years./outcome-earthquake-forecasting-xgboostAI for Energy and Sustainability – Shell ProjectPoster · PBL, 2025170https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedcfb783092fef11884f9e_thumb-v3b-earthquake.jpg
An Innovation Engine for a Large BankOutcomeIndustryAI + Business & FinanceA team used JPMorgan as a case study, diagnosed why innovation stalls in big banks, and proposed a Startup Factory plus a shared platform, modeled at about $256M NPV on their own assumptions./outcome-bank-innovation-engineMachine Learning in Quantitative Finance – J.P. Morgan ProjectPresentation · PBL, 2025160https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3fafcac998bd7f1a408_thumb-v3-jpm-innovation.jpg
Biomedical Search by Meaning, Not KeywordsOutcomeResearchAI + Vision & LanguageA team trained a teacher–student model to match biomedical questions to answers by meaning. Their adjusted training stayed stable longer, and they reported a data leak in their own splits./outcome-biomedical-semantic-similarityAI in Natural Language Processing – Hugging Face ProjectPoster · PBL, 2025150https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3fa4e0c7ee9f506d042_thumb-v3-biomedical-sts.jpg
Visual SLAM Built from ScratchOutcomeResearchAI + Physical IntelligenceA team built graph-based visual SLAM from scratch and tested it three ways: on KITTI, on a marker photo set and live with an RGB-D camera, reporting where it broke./outcome-visual-slam-from-scratchAI and Robotics for Mobile Robot Manipulation – Boston Dynamics ProjectPresentation · On-Campus Experience, August 2025140https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3fa6e361e0c0967ccc4_thumb-v3-visual-slam.jpg
Smart Glasses Design for Brain Tumor MRIOutcomeIndustryAI + Physical IntelligenceA team designed AI smart glasses that detect brain tumors on MRI and describe them in plain language. The detection pipeline ran on a smartphone; the glasses design came to 68 g and 9.6 W./outcome-smart-glasses-brain-tumor-mriAI in HardwarePoster · PBL, 2025–26130https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3f96e361e0c0967ccb1_thumb-v3-smart-glasses.jpg
When a GAN Makes a Medical Classifier WorseOutcomeResearchAI + Health TechA student compared five pneumonia-detection setups. Simple augmentation scored best (0.993 AUROC); CycleGAN images lowered every metric, and he traced why./outcome-pneumonia-augmentation-vs-cycleganAI and Computer Vision in BioTech – Novo Nordisk ProjectPresentation · PBL, 2025120https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3f9fcac998bd7f1a3a5_thumb-v3-pneumonia.jpg
Stereo Depth: Classic Matching vs. Deep LearningOutcomeIndustryAI + Physical IntelligenceA team built a stereo pipeline from calibration to 3D point clouds and compared classic matching with deep learning: denser depth from the network, more reliable object depth from geometry./outcome-stereo-depth-classical-vs-deep-learningComputer Vision and Robotics – Amazon Robotics ProjectPoster · PBL, 2025110https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3f936debf68379b0333_thumb-v3-stereo-depth.jpg
One Walkable Space from Many 3D ScansOutcomeResearchAI + Vision & LanguageA team stitched separately trained 3D Gaussian Splatting scenes into one space you can move through, with a trajectory map, portals between scenes and placed 3D models./outcome-interactive-3d-gaussian-splattingComputer 3D Graphics and Deep Learning – NVIDIA ProjectPoster · PBL, 2025100https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedc3f91d11e9c07c3388ef_thumb-v3-3dgs-interactive.jpg
3D Perception from a Single CameraOutcomeIndustryAI + Physical IntelligenceWith one ordinary camera, a team combined camera tracking, 3D reconstruction and object detection into one pipeline, checking each stage against ground truth./outcome-single-camera-3d-perceptionAI and Robotics for Mobile Robot Manipulation – Boston Dynamics ProjectPresentation · PBL, Nov 202590https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab6cc97880185526ec0_thumb-v3-single-camera.jpg
Why AI Jobs Slow Down When GPUs Are Not FullOutcomeIndustryAI + Business & FinanceUsing Alibaba GenTD26 cluster traces, a team built linked Tableau dashboards to find out why AI jobs were delayed. In 12 of 14 high-delay windows, average GPU use was low./outcome-gpu-cluster-delay-diagnosisAI and Data Visualization for Social Impact – Tableau ProjectDashboard · On-Campus Experience, Aug 202680https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab6c51c3f3a7a236374_thumb-v3-gpu-delay.jpg
Knee MRI: Fix the Data Before the ModelOutcomeResearchAI + Health TechWith 4,407 knee MRI studies and only 58 labeled by radiologists, a team found the problems hidden in the data before training anything, then used LLMs to turn reports into labels (0.80–0.89 AUROC)./outcome-knee-mri-data-labelingAI and Computer Vision in BioTech – Novo Nordisk ProjectPresentation · On-Campus Experience, Aug 202670https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab6e46d0e01cfd189d0_thumb-v3-knee-mri.jpg
News Sentiment and the DollarOutcomeResearchAI + Business & FinanceA team used an LLM to score risk sentiment in 38,384 financial news articles. A one-point US–euro area sentiment gap predicted an 8.1 basis-point next-day move in the dollar./outcome-news-sentiment-currencyAI and Economics: Decision-Making with Language Models – Bloomberg ProjectResearch paper · PBL, Aug 202660https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab6e28fe8f3c0147246_thumb-v3-news-sentiment.jpg
Glare Suppression Before 3D ReconstructionOutcomeResearchAI + Vision & LanguageA team cut glare from photos before 3D reconstruction, correcting only where the reflection showed up across several views. Across 561 images, fine detail held (0.995–0.999 correlation)./outcome-glare-suppression-3d-gaussian-splattingComputer 3D Graphics and Deep Learning – NVIDIA ProjectResearch paper · PBL, Aug 202650https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab6e982df1aee3a4ac9_thumb-v3-glare-suppression.jpg
Demand Forecasting to Inventory DecisionsOutcomeIndustryAI + Business & FinanceA team fed demand forecasts from 112,634 e-commerce order lines into a two-stage inventory optimizer. In simulation, cost fell 2.2–3.8% at a 94% fill rate./outcome-ecommerce-demand-inventoryAI in Operations Strategy – Amazon ProjectPoster · On-Campus Experience, Aug 202640https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab5e982df1aee3a4a70_thumb-v3-demand-inventory.jpg
Scoring Safe Cellular RejuvenationOutcomeResearchAI + Health TechA team scored youth, cell identity and risk separately in single-cell data, and found a foundation model's aging signal mostly tracked sequencing depth./outcome-safe-cellular-rejuvenationAI for Life Sciences Innovation – Genentech ProjectResearch paper · PBL, Jul 202630https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbf1f236d3a95a97c84cc_thumb-v3c-cellular-rejuvenation.jpg
Smarter Memory Scheduling for AI ChipsOutcomeResearchAI + Physical IntelligenceA team built its own simulator for an in-car vision network. The chip sat idle 87% of the time waiting on memory; reusing activations made it 2.1x faster, in simulation./outcome-mobilenetv2-memory-prefetchingAI in HardwareResearch paper · PBL, Jun 202620https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbf1f4e0c7ee9f505c6fe_thumb-v3c-memory-prefetching.jpg
Nam PhamStudentSix industry projects across robotics, hardware and strategy, two of them on campus in Boston-Cambridge./portfolio-nam-phamUniversity of Tokyo40
Rumaisa KashifStudentThree industry projects, then co-founder of the AI+X Global Talent Community./portfolio-rumaisa-kashifNational University of Singapore20
Bowen LiuStudentFrom first-year student to AI oncology intern in ten months. Eight industry projects and the evaluations behind them./portfolio-bowen-liuSouth China University of Technology10
Brain Tumor Segmentation and VolumetryOutcomeIndustryAI + Health TechA team adapted a pretrained CNN–ViT model to brain MRI and shipped a web tool that outlines tumors and measures their volume. Dice 0.848 on 73 held-out patients./outcome-brain-tumor-segmentationAI and Computer Vision in BioTech - Novo Nordisk ProjectWorking web tool · On-Campus Experience, Aug 202610https://cdn.prod.website-files.com/6842c25bada7cd119b9732fb/6aedbab3d279b10a2462dc32_thumb-v3-brain-tumor.jpg








