Fellows, and where their work landed.
Every study below is a real student with a verifiable result: a published paper, a competition title, or an admission letter. Read how each one happened.
Sadaf ShireenFrom no research experience to IEEE published and admitted to Brown
Sadaf joined the YRI Fellowship the summer before senior year with an interest in AI but no research background, and was paired with a PhD mentor. In 10 weeks she completed original research on neural network optimization for real-time classification systems, and the paper was accepted at an IEEE conference. The publication became the centerpiece of her college applications. The research did not stop at the paper: she went on to build a venture on top of the same work, and in April 2026 it was selected for the WUHC Incubator Pitch Competition at Wharton.
Mubashir SuhailFrom zero research experience to ISEF 2026 Qualifier
Mubashir tackled a structural barrier in astrophysics: gravitational-wave machine learning normally needs $50,000 or more per year in GPU infrastructure. He trained a lightweight CNN of roughly 92,000 parameters on LIGO-Virgo detector strain data using free Google Colab, then engineered a peer-to-peer distributed training framework so that 100 standard laptops can match a single NVIDIA V100 GPU. The work qualified him for ISEF 2026.
Aditya SinglaFrom zero research experience to Best Oral Presentation at IEEE in Tokyo
Working with his YRI PhD mentor, Aditya built a two-step neural network pipeline for anomaly detection in large-scale cloud infrastructure, trained on IBM Cloud Console telemetry covering 39,365 data points and more than 117,000 features. His system detected 16 of 25 anomalies against 6 of 25 in prior research, a 2.7x improvement. He presented at IEEE AAIML 2026 in Tokyo, March 20 to 22, 2026, and won Best Oral Presentation in his session.
Aaryan SenthilvananThree papers accepted at three international conferences in three countries
Aaryan produced three distinct research projects, each addressing a different open problem in computational biology. InteractionFormer simulates the human interactome using SE(3) equivariant graph neural networks with anisotropic network models, GPCRome maps the complete human GPCR signaling connectome to deorphanize receptor agonism, and ORACLE designs a phase-state coacervate nanocarrier system for epigenetic reprogramming of cancer cells. All three were independently peer reviewed and accepted, and he presents in Washington DC, Italy and Spain.
Ansh KumarPublished at MIT URTC and built a clinical-grade cancer detection model at 15
Ansh produced two major research projects at once. His Parkinson's disease work on biomechanical interventions for improving motor outcomes was accepted at MIT URTC, an IEEE-affiliated conference. His second project trained a gradient boosting classifier on 7,240 TCGA samples using five DNA methylation biomarkers, reaching 98.2% sensitivity and an AUC of 0.983 for non-invasive lung cancer detection.
Devansh GuptaIEEE published at 14 and founder of a 7-figure health-AI company
Devansh built a framework that analyzes ECG microstate instability patterns as objective biomarkers for Major Depressive Disorder, a condition affecting more than 280 million people that is still diagnosed largely through subjective clinical assessment. His research combined advanced signal processing with machine learning and connected cardiology, neuroscience, immunology and machine learning. He was one of the youngest first authors at the conference at just 14 years old.
Ashwajit WarwatkarFrom enrollment to a Springer paper and an international conference talk in 2 months
Ashwajit developed the Replay-Gated Cascade Consolidation model, a 1,000-neuron Izhikevich spiking network with spike-timing-dependent plasticity and Fusi-type slow cascade weights. He stated three predictions in advance and confirmed all three across more than 700 simulation runs, scaling up to 5,000 neurons. Removing replay dropped retention from 0.286 to 0.037, and the harmonic-series prediction fit with an R squared of 0.828. He presented the work himself at BICA 2026 in Merida, Mexico, a 15-minute talk followed by 5 minutes of questions from the room. YRI then helped him take the next step: entering the 2026 Breakthrough Junior Challenge, where he made a short film explaining the Kakeya conjecture, a famous problem about how little space a shape needs to point in every direction. His second paper, on CT-derived texture features that stratify survival in Stage III non-small cell lung cancer, was accepted for a poster presentation at SPIE Medical Imaging 2027 in Vancouver.
Elliot WongA personalized helmet design study, accepted at an international healthcare robotics symposium
Elliot came to YRI as a 10th grader with a question that had real clinical stakes: can protective helmets be personalized to the actual geometry of a person's head and brain? He had tinkered with a custom-fit prototype on his own, but turning it into research required rigor he did not yet have. With his mentor, the idea became an MRI-guided approach to personalized helmet liner design for head-injury prevention, using 3D head scanning and subject-similar MRI templates to tailor the liner to the individual rather than a one-size average. The paper was accepted at the 2nd International Symposium on BioMechatronics and Robotics in Healthcare (BioMRH 2026) in Glasgow, UK, with proceedings published in Springer's Lecture Notes in Networks and Systems series.
Karan AggarwalA high schooler's econometrics paper, accepted at an international finance symposium
Karan joined the YRI Fellowship in March 2026 wanting to do real work in finance, and was paired with a mentor to design a study most students would not attempt before graduate school: a Difference-in-Differences analysis of how short-duration geopolitical shocks move sector stocks relative to the broader market, using the 2016 Uri attack and 2025 Operation Sindoor as events, with InterGlobe Aviation against the BSE Sensex. Six months after enrolling, the full paper was accepted at the World Finance & Banking Symposium, an international academic finance conference held in Riga, Latvia, where he will present alongside university researchers in December 2026.
Nidhi KulkarniFrom exploring research ideas to first-author IEEE publication in one summer
Nidhi arrived with two separate research interests and her mentor helped her combine them into one project. She built an AI-driven multi-omics framework integrating gut and skin microbiome data to identify non-invasive biomarkers for PCOS and autoimmune disorders, using FastQC, Trimmomatic, Bowtie2, Kraken2 and Bracken pipelines with Random Forest, XGBoost and LightGBM models. The paper was published at IEEE ICCCA 2025.
Ruthwik DhamaFrom exoplanet enthusiast to creator of a novel habitability framework
Ruthwik expanded his initial idea from analyzing a single planet into building an entirely new habitability index. METHI uses binary classification, unsupervised clustering and ensemble-based regression to improve on fixed-heuristic indices like ESI, PHI and SEPHI, achieving a 0.903 score and identifying the top 10 habitable exoplanet candidates. He also built a public web interface so anyone can input a planet name and retrieve a real-time habitability score.
Aarav BrahmbhattFrom zero research experience to IEEE-published author as a 9th grader
Aarav built an AI framework that predicts traffic incident likelihood, expected waiting time and CO2 emissions from static traffic scene images. YOLO object detection counts and classifies vehicles, and those features feed a Random Forest regressor for multi-target prediction. The paper was published in IEEE Xplore through ICITSIF 2026 in Indore, India, May 29 to 30, 2026.
Raeyaan MuppaneniBuilt a low-cost bionic arm with 93% real-time accuracy and published it at IEEE
Raeyaan built a complete system from scratch: a low-cost bionic robotic arm that reads EMG muscle signals from the forearm, classifies gestures with AI and drives servos in real time. MyoWare sensors feed an ESP32 that transmits over MQTT to a Raspberry Pi 4 running inference. He compared KNN, Random Forest, MLP and SVM across three feature pipelines, and Random Forest on raw EMG envelopes reached 93% real-time accuracy.
Suriya Dev SaravanakumarAccepted at the top biomedical engineering conference as a 9th grader
Suriya built a machine learning framework that analyzes eye-tracking measures including fixation duration, saccade amplitude and blink rate to separate Alzheimer's patients from healthy controls. He tested Random Forest, XGBoost, a feed-forward neural network and ensemble methods with ten-fold cross-validation under signal degradation, reaching a ROC-AUC of 0.75 with horizontal eye-position measures as the strongest indicators. He also built a Streamlit prototype for real-time eye-tracking visualization.
Arham SethiBuilt the first framework to map pre-emergence weed risk, published at IEEE
Arham developed WERM, a framework that estimates pre-emergence weed risk by fusing soil moisture, residue cover, structural disturbance and growing degree days. It combines ResNet-18 moisture and residue estimation from RGB field images, GLCM texture analysis and NASA POWER API thermal data through a log-space multiplicative model. Validated on 333 field images and 30 ground-truth soil samples with farmers of 20 to 30 years experience, it reached a ground-truth correlation of 0.879.
Aly DhedhiDesigned brain-cancer peptide therapeutics and published at IEEE as a 10th grader
Working with a YRI PhD mentor, Aly designed and evaluated four modular peptide constructs fusing the blood-brain barrier shuttle motif Angiopep-2 with the EGFR-targeting sequence GE11, aimed at the tumour-exclusive receptor variant EGFRvIII in glioblastoma. The pipeline used ESMFold for structure prediction, PatchDock for docking and BioPython for physicochemical profiling. Construct P3, a cyclized GE11 variant, was the only candidate rated High for blood-brain barrier permeability.
Akul NehraBuilt a regulatory-compliant federated AI framework and published at IEEE at 15
Akul built a federated ensemble learning framework combining five heterogeneous base learners across 24 federated clients, using differential privacy with Renyi DP accounting, meta-learning aggregation via stacking and secure proxy training. Trained on 150,000 real-world GMSC loan samples plus 45,000 synthetic samples with non-IID heterogeneity, it reached 91.8% accuracy, just 1.7 percentage points off centralized performance. The framework simultaneously satisfies GDPR Article 32, ECOA anti-discrimination requirements and Basel III calibration standards.
Atharv VedBuilt a CRISPR target prediction pipeline for heart disease, accepted at two IEEE conferences
Atharv developed a computational framework for predicting safe, high-specificity CRISPR-Cas9 guide RNAs targeting the cardiac remodelling genes MYH7, ACTC1, TNNT2 and NPPA. He retrieved sequences from NCBI GenBank, scanned for SpCas9 PAM sites using Biopython and scored candidate guides through BLAST-based off-target profiling with mismatch heatmapping and specificity scoring. All three tested guides hit roughly 100% editing efficiency, but only gRNA2 targeting MYH7 combined that with the lowest off-target burden.
Ryan VoFrom no coding experience to IEEE published with AI-optimized skin grafts
Working with a YRI PhD mentor, Ryan built an AI framework integrating Multivariate Normal Distribution and polynomial regression with a 3D Generative Adversarial Network to optimize scaffold designs for bioprinted skin grafts. The model trained on synthetic datasets derived from the CO2Wounds-V2 chronic wounds dataset from leprosy patients. It reached an R squared of 0.59 and identified optimized parameters for porosity, pore size, thickness and biomaterial selection.
Sai TarekarFrom zero coding experience to an IEEE acceptance as a 9th grader
Sai developed AHIRRS, an Automated Hybrid In-Situ River Remediation System for cleaning the Yamuna River. The 22km Wazirabad-Okhla stretch in Delhi-NCR receives roughly 3,491 million liters of sewage daily against a treatment gap of 900 million liters per day. His framework proposes compact decentralized treatment units placed directly in the river at pollution hotspots, using machine learning to analyze water quality data and optimize treatment in real time.
Aarnav BhatFrom aspiring AI researcher to IEEE acceptance with AUC 0.983
Aarnav developed a machine learning approach for discovering biomarkers in lung cancer, analyzing genomic data to identify signatures associated with the disease. His model achieved an AUC of 0.983, near-perfect discrimination, pointing toward earlier detection when treatment is most effective. The paper was accepted at IEEE ICITSIF 2026.
Sai PasuparthiFrom curious about COVID evolution to an IEEE conference acceptance
Sai conducted a genomic analysis of Omicron variant mutations, combining epidemiological analysis with evolutionary insight to understand how the virus evades immune responses. Using computational phylogenetics and genomic surveillance, she systematically cataloged spike protein mutations across lineages and identified 132 immune escape mutations. The findings inform vaccine development and pandemic preparedness.
Neil VoorePublished medical AI research on evaluation bias as a 9th grader
Neil examined how standard evaluation protocols inflate reported accuracy in nail abnormality detection models. He showed that random data splits allow leakage between training and test sets, and that models with near-perfect accuracy under random splits decline sharply when evaluated across different data sources. Region-of-interest training focused on the nail plate improved external F1-score and reduced background bias.
Ayaan RustagiApplied astronomy techniques to detect Alzheimer's at 88.9% accuracy
Ayaan built a cross-domain approach that adapts astrobiology signal processing, including pulsar timing techniques used to pull faint signals out of noisy astronomical data, to detect Alzheimer's patterns in brain scans. His machine learning models trained on those cross-domain features reached 88.9% accuracy. The work was published in IEEE and won 3rd place at his science fair.
Shaswat SenthilkumarFrom curious 10th grader to 1st place at a Princeton University science fair
Shaswat built a machine learning framework that predicts stem cell differentiation outcomes from RNA sequencing data. His pipeline processes gene expression data through feature extraction and ML classification to predict cell fate, helping researchers see which genes matter most in differentiation decisions. The research has implications for regenerative medicine and treating diseases like Parkinson's, diabetes and heart disease.
Avyay G.From curious 9th grader to 1st place science fair winner and state qualifier
Avyay built a multi-model AI pipeline predicting time-to-onset of respiratory disease from combined air pollution exposure and genetic susceptibility, using WHO Global Air Quality data with GEO gene expression datasets. He compared Kaplan-Meier, Cox Proportional Hazards, AFT, Random Survival Forests and DeepSurv with SHAP explainability for clinically actionable predictions. Higher pollution accelerated disease onset, especially in genetically susceptible individuals.
Aarush MahajanSelected as 1 of 14 global finalists from over 8,000 participants
Aarush designed the Carbon Conversations & Community Action Toolkit, providing frameworks for measuring carbon footprints, facilitating community discussions and implementing sustainability initiatives. Over 5 months he refined his Climate Action Project through mentorship and hands-on experimentation and built a dedicated project website. He was selected as 1 of 14 global finalists from more than 8,000 participants and represented India at the PGC 2025 Final Summit in San Francisco.
Aditya ArvindReading movement straight from the brain, for the patients whose muscles cannot be read
Prosthetic hands are usually controlled by electromyography, which reads electrical activity in muscle. For conditions like Spinal Muscular Atrophy that is the one signal that is not there to read. Aditya built the alternative: a pipeline that classifies motor imagery straight from EEG, bypassing the muscle entirely. He benchmarked three deep learning architectures against three classical machine learning models on the PhysioNet Motor Movement/Imagery dataset, separating rest from imagined left-hand and right-hand movement, and ran every model twice, once on the five cleanest subjects and once across all 109, to separate raw accuracy from real generalization.
Abhigna NaikA low-cost AI screening system for rural clinics, accepted at an IEEE conference
Over 65% of India's population lacks timely access to diagnostic imaging. Abhigna set out to close that gap with ArogyaScan, a point-of-care screening framework that community health workers without specialist training can run. It screens pregnancy, liver, kidney and gallbladder conditions from ultrasound and lung conditions from chest X-rays, using organ-specific deep learning classifiers. The best model reached 98.08% accuracy on lung classification, held up on external validation, and the full system costs 70 to 80% less than a conventional hospital ultrasound setup while running fully offline.
Amaay GuptaA drone that chooses its own emergency landing, accepted at an IEEE conference
Detecting that a drone is in trouble is not the same as knowing what it should do next. Amaay turned his idea into Skyshield, a decision layer that takes motion-sensor evidence and the vehicle's current state and picks one of seven emergency actions. It first rules out actions the drone physically cannot perform, then ranks the rest by modeled consequences. He evaluated it at serious scale: 4.2 million paired decisions, 209 full-flight replays, 450 matched PX4 and Gazebo simulation runs, and a replay of physical sensor recordings.
Arjun Vinayaka DossAI-designed heat shields for reusable spacecraft, accepted at an IEEE conference
Every spacecraft that returns to Earth depends on a heat shield surviving re-entry. Arjun used machine learning to predict how fibre-reinforced composites perform under extreme heat, drawing on NASA's TPSX materials data covering density, thermal conductivity, heat capacity and a thermal shielding efficiency index. Random Forest and neural network models identified LI-900 and carbon-phenolic as the strongest combinations for insulation, pointing to a faster way to select and design materials for reusable spacecraft.
Ben ZhangPredicting ACL injuries and recovery time in athletes, accepted at an IEEE conference
ACL tears are among the most common serious injuries in sport, and most research looks at anatomy while ignoring what athletes actually do in training. Ben built interpretable machine learning models, including Random Forest and Gradient Boosting, on basketball athlete performance data such as training hours and intensity. His work predicts not just the risk of a first-time ACL injury but also how long recovery will take, a question most existing studies treat as a fixed input rather than something to forecast.
Nishanth AlampallyDiscovered the Jitter Paradox in physics-informed AI, accepted at an IEEE conference
Physics-informed neural networks learn solutions that obey physical laws, and newer architectures called Kolmogorov-Arnold Networks often solve equations more accurately. Nishanth asked a sharper question: are those solutions good enough to rediscover the underlying equation? Comparing both architectures on the heat equation under noise, he found a Jitter Paradox. The more accurate networks produced rougher derivatives that corrupted symbolic equation discovery, while simpler networks recovered the physics more reliably. He also proposed a new metric, the Bias Inheritance Ratio, to measure how model errors leak into discovered equations.
Raaghav Sai GundaboluModeling what breaks photonic quantum computers, accepted at an IEEE conference
Photonic quantum computers are promising, but every imperfect photon source, lossy channel and noisy detector eats into their performance. Raaghav built a simulation framework that models these losses in small fusion-based quantum benchmarks, pairing an analytical model with Monte Carlo validation. His predicted success rates of 25.5%, 20.7% and 14.8% across three benchmark circuits matched the simulations within 95% confidence intervals, and his analysis shows exactly which component limits each circuit.
Rishik AsatiA 9th grader stopped AI agents from leaking secrets, accepted at an IEEE conference
AI agents that read emails, documents and websites can be tricked by hidden instructions into leaking private data, an attack called indirect prompt injection. Rishik built a lightweight defense that separates trusted instructions from untrusted content, detects instruction-like attacks, gates sensitive outputs and tracks risk across a conversation. Across 150 test cases and seven attack categories, active detection cut observed leakage of protected secrets to zero, while passive measures like warnings alone did nothing. He also documented a flaw in his own evaluation openly rather than hiding it, exactly the rigor reviewers look for.
Saanvi KumarFinding which customers a message actually persuades, accepted at an IEEE conference
Most campaigns send everyone the same message and measure the average effect, which hides the people who respond strongly and the ones who are put off. Saanvi built an uplift modeling framework using X-Learner models to estimate each individual's response to different interventions and choose the best one per person. Personalized targeting produced $6.50 more incremental spending per customer than a uniform approach, and Qini curve analysis revealed a concentrated group of persuadable customers driving most of the gain.
Sidharth S. IyerBuilt a pipeline to study disease in sea stars, accepted at an IEEE conference
Sea star populations have been devastated by disease, and the clues often hide in genetic data left over from other studies. Sidharth built a full computational pipeline for the bat star that strips out the animal's own DNA, classifies the remaining microbial sequences in the cloud, and removes false positives with strict similarity and alignment filters. Custom R tools then turn the results into heatmaps across locations and time. He turned research he had been carrying on his own into his first accepted paper.
Ishaan MenonTurned waste tires into cleaner fuel, published at an IEEE conference
Billions of discarded tires pile up because they barely degrade and are hazardous to burn. Ishaan studied pyrolysis, which breaks tires down into fuel, gas and carbon, and designed a hybrid catalytic system pairing a bimetal-modified zeolite with catalysts made from industrial waste. He analyzed reaction kinetics, activation energies and product distributions, built predictive models to optimize yields of aromatic hydrocarbons and low-sulfur fuels, and assessed catalyst regeneration and techno-economic feasibility. The work was published at an IEEE conference.
Arfa HaseebUsing AI to predict how cancers respond to drugs, accepted at an IEEE conference
Two patients with the same cancer can respond completely differently to the same drug, and much of that difference is written in their biology. Arfa built machine learning models that combine multiple layers of biological data, known as multi-omics, to predict how cancers will respond to specific drugs. The goal is to help match patients to the treatments most likely to work for them. Her paper was accepted at an IEEE conference.
Mohammed AlnuwaiserMaking AI-planned drone teams fail safely, accepted at an IEEE conference
Language models are starting to plan missions for teams of drones, but a single bad plan can cascade across the whole fleet. Mohammed studied where validation checks should sit in that planning process, and how their placement affects both how well failures are contained and how much compute the system spends. His paired study gives a practical answer to a question that matters as AI takes on more control of physical systems.
Diya AnneBuilt an AI digital twin of the patient, published at an IEEE conference
Digital twins, live virtual copies of physical systems, have long been used in factories and engines. Diya brought the idea to medicine. Her framework builds an evolving virtual replica of a patient from wearable sensors and electronic health records, then uses deep sequence modeling, multimodal data fusion and reinforcement learning to forecast adverse health events, simulate personalized treatment pathways and refine care recommendations from real-world feedback. Evaluated on the MIMIC-III clinical dataset and a real-time pilot study, it improved prediction accuracy, alert latency and intervention efficacy over existing models.
Harleen BadwalAn AI model that flags endometriosis risk without surgery, accepted at an IEEE conference
Endometriosis affects about 10% of women of reproductive age, and diagnosis often takes years because it usually requires surgery to confirm. Harleen built EndoRisk-Ensemble, a model that estimates risk from symptoms and clinical history alone. She found that a conventional model scored a high AUC of 0.959 but caught none of the high-risk patients, then fixed it with minority-class oversampling, a three-model ensemble and confidence-weighted voting. Her model reached an AUC of 0.990 and identified 4 of 5 high-risk patients in the held-out test set.
Krishik SinghA 9th grader predicted hidden satellite collision risks, accepted at an IEEE conference
Satellite operators receive repeated warnings about possible collisions, and the dangerous cases are the ones that look safe early and turn risky later. Krishik formalized this hidden-escalation problem and trained models on the trajectory of risk messages, using only data available two days before closest approach. Across 10,460 real encounter events, he found that 25% of the eventually high-risk cases were invisible to standard threshold rules at that point, and his trajectory features lifted ROC-AUC on that hard group from 0.885 to 0.915.
Naitik GuptaAI that screens materials for cleaning heavy metals from water, accepted at an IEEE conference
Metal-organic frameworks can pull toxic heavy metals out of water, but testing each one takes long series of experiments. Naitik compiled 289 experiments from 49 studies and trained an optimized CatBoost model to predict how much metal each material can capture. It reached an R-squared of 0.897, and SHAP analysis showed which properties matter most, matching established adsorption theory, so researchers can screen candidates in software before the lab.
Khyati SwainAI that predicts immune reactions to heart implants, accepted at an IEEE conference
Stents, valves and other cardiovascular implants save lives, but some patients' immune systems react against them. Khyati built an AI-driven risk predictor that estimates how a patient's immune system is likely to respond to a cardiovascular implant, aiming to flag high-risk patients before the procedure rather than after complications appear.
Melissa NnamA 9th grader built an explainable anemia screening tool, accepted at an IEEE conference
Anemia affects billions of people and is often caught late. Melissa built an explainable screening tool that estimates anemia severity from standard complete blood count features, and shows which blood measurements drive each prediction so clinicians can trust and check the result.
Svanika Naidu DoddavarapuTaught small AI models to track what they believe, accepted at an IEEE conference
Small language models can reach a right answer for the wrong reasons. Svanika studied whether compact instruction-tuned models reason better when they explicitly track their belief state step by step, looking beyond final answers to how the model gets there.
Devansh DekaEstimating blood sugar without a needle, accepted at an IEEE conference
Millions of people with diabetes prick their fingers every day to check blood sugar. Devansh worked on a needle-free alternative: turning the optical pulse signal measured by a simple sensor into time-frequency images with derivative-synchrosqueezed transforms, and using them to estimate blood glucose non-invasively.
Tarek OsamaFrom Egypt to an IEEE acceptance on climate-resilient rice
Soil salinization is one of the biggest threats to rice, the staple food for more than half the world. Tarek built OryzaGen, a reproducible pipeline that combines high-throughput transcriptomic profiling, public 3K Rice genomic phenotypes, published GWAS loci and population-blocked machine learning classifiers. The pipeline isolated 28 high-confidence candidate genes within 100 kb of established salinity GWAS peaks, including LOC4329464, LOC4329812 and LOC4352160, giving breeders a transparent shortlist for marker-assisted selection and targeted crop engineering in climate-vulnerable regions.
Teaching AI to know when it is wrong about cancer immunotherapy, accepted at an IEEE conference
Immunotherapy works brilliantly for some cancer patients and not at all for others, and a model that predicts response is only safe if it also knows when it is likely to be wrong. Akhilesh evaluated exactly that across two real melanoma cohorts, 110 patients on CTLA-4 blockade and 143 on PD-1 blockade. He benchmarked three model families with leakage-free repeated cross-validation, measured calibration and patient-level uncertainty with bootstrap ensembles, and tested whether selective prediction and conformal prediction could flag unreliable calls. The result is an honest, rigorous answer to a question clinicians actually ask.
Making graph neural networks safe for hard optimization problems, accepted at an IEEE conference
Graph neural networks can score which nodes matter in an optimization problem, but their raw outputs often break the rules the solution has to follow. Manomay studied this on minimum vertex cover, a classic hard problem, and found raw network outputs were valid in only 38.67% of cases. His pipeline repairs every uncovered edge and then prunes redundant vertices, restoring 100% validity and cutting the solution from 24.87 to 21.90 vertices, with a mean optimality gap of 2.75% against exact solutions.
An adaptive digital twin for smarter machining, accepted at an IEEE conference
CNC machines cut the precision parts that modern manufacturing depends on, and tuning them means balancing speed, quality and cost at the same time. Adit built an adaptive digital twin, a virtual model of the machining process that adjusts to each component's complexity, and used machine learning to optimize across multiple objectives. Reviewers noted the framework outperformed fixed, rules-based and standard machine learning approaches, with clear analysis of noise robustness and ablations.
Used NLP to audit the research on GLP-1 drugs and binge eating, accepted at an IEEE conference
Research on GLP-1 drugs, the class behind today's weight-loss medications, is exploding faster than anyone can read it. Athena asked whether a transparent language-processing pipeline could sort that literature the way expert reviewers do. Working with 1,422 records from a curated evidence map, she built a classifier that labels findings on brain measures, craving, binge outcomes, appetite and metabolic outcomes, using study-grouped cross-validation so no study leaked between training and testing. Her model reached an F1 score of 0.894 on appetite and satiety outcomes.
Tested how well telescopes catch the shortest exoplanet transits, accepted at an IEEE conference
Space telescopes find planets by catching the tiny dip in starlight when a planet crosses its star, and very short transits are easy to miss depending on how often the telescope samples. Ana built a reproducible injection-and-recovery study modeled on TESS-like observations, planting simulated transits into data and measuring which ones survive detection, to map the risk that observing cadence hides real planets.
Asked whether two nearby pulsars explain a cosmic-ray mystery, accepted at an IEEE conference
Detectors in space see more high-energy positrons than standard physics predicts, and nearby pulsars are a leading suspect. Yuvan built a simplified two-zone diffusion model to test whether two nearby pulsars, Geminga and Monogem, could contribute to this positron excess, tackling an open question in astroparticle physics.
Decoded genetic variants behind a hidden heart disease, accepted at an IEEE conference
Arrhythmogenic cardiomyopathy is an inherited heart disease that can cause sudden cardiac death in young, healthy people. Ajwa studied missense variants in the five desmosomal genes linked to it, interpreting each variant by where it sits on the protein interfaces that hold heart cells together, to help tell disease-causing mutations from harmless ones.
Compared devices that turn vibration into electricity, accepted at an IEEE conference
Everyday vibrations from machines, vehicles and footsteps can be turned into electricity to power small sensors. Ariham compared three harvester designs, passive linear, symmetric bistable and tunably asymmetric, under representative real-world excitations to find which designs capture the most energy in practice.