# Vaibhav Balloli: full site text > Vaibhav Balloli, applied AI researcher and CS Ph.D. candidate at Michigan, building interactive AI systems for health, conservation, and education. Generated 2026-10-24T03:36:07.956Z from https://profile.vballoli.com/. Short index: https://profile.vballoli.com/llms.txt · Structured data: https://profile.vballoli.com/agents/ ## Profile Vaibhav Balloli is an Applied AI Researcher and Computer Science Ph.D. Candidate at University of Michigan, advised by Elizabeth Bondi-Kelly. I study how to make deployed AI systems reliable across long conversations, shifting user intent, and real-world constraints. My work spans agent post-training, evaluations designed with domain experts, instruction following, and test-time monitoring, with applications in reproductive health, wildlife conservation, and road safety. Research areas: Large Language Models, Reinforcement Learning, Representation Learning, Human-AI Interaction, Agentic Systems, Machine Learning. Contact: balloli.vb@gmail.com. ## Research lineage - Helping experts steer retrieval. For conservation biologists & rangers. Are They the Same Picture? Adapting Concept Bottleneck Models for Human-AI Collaboration in Image Retrieval (Let experts intervene on concepts mid-search) -> SEEK-CBM: Editable and Interpretable Retrieval for Elephant Re-Identification (Modern backbones, 497% better retrieval). Stage: Research + Deployment. Deployment: To be deployed for the elephant work. - Deploying trustworthy health AI systems. For birthing people, clinicians & health workers. "Where is this coming from?" Uncovering Trustworthiness Ideals in AI-powered Peripartum Information Seeking (Ask them first: four focus groups, three stakeholder groups) -> PATHFinder Agent for Tailored Prenatal Care (Build to that spec: a care-planning agent) and RELIANCE: Curating and Evaluating Reproductive Health Information on Social Media (Then measure the ecosystem it lives in). Stage: Research + Deployment. Deployment: TBD in Michigan. - Selecting best system configuration. For adaptive systems. Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception (Trade accuracy for latency as the scene changes) -> Bridging the Gap: Dynamic Learning Strategies for Improving Multilingual Performance in LLMs (Same idea, now choosing prompts and models per query). Stage: Research. ## Systems and deployments ### HAMS: automated driver licence testing Microsoft Research India, 2021 – 2022. Status: Deployed. Societal Impact through Cloud and AI Fellow. Worked on the computer vision, bandit, and localisation components as the system scaled. Smartphone-based automated licence testing deployed across India, serving 440,000+ candidates and replacing a subjective manual assessment with an auditable one. Areas: Computer vision, On-device inference, Bandits, Localisation. Links: Project https://www.microsoft.com/en-us/research/project/hams/ · Deployment https://www.microsoft.com/en-us/research/project/hams/automated-driver-license-testing/ · Press https://news.microsoft.com/en-in/features/microsoft-ai-automates-drivers-license-test-india/ ### PATHFinder: prenatal care planning agent University of Michigan, 2025 – 2026. Status: In progress. Lead author. Designed and built the four-stage agent: patient intake, dynamic interaction, plan synthesis, and clinician oversight. Conversational agent that produces individualised prenatal care plans aligned to ACOG's PATH guidelines and surfaces local resources. Frontier models scored against expert rubrics across five clinical dimensions, with clinician evaluation in Michigan health systems underway. Areas: LLM agents, Long-context optimisation, Clinical rubrics. Links: Project https://realize-lab.github.io/PATHFinder · arXiv https://arxiv.org/abs/2607.24768 ### CHAIR & SEEK-CBM: steerable wildlife re-identification University of Michigan · MIT, 2023 – 2026. Status: In progress. Lead author on CHAIR; co-author on the SEEK-CBM follow-up with the Beery Lab at MIT. Concept-bottleneck retrieval that lets conservation experts intervene mid-search, built for platforms like ElephantBook and evaluated on elephant populations in the Masai Mara. SEEK-CBM's modern backbones improved retrieval performance by 497% over CHAIR. Areas: Concept bottleneck models, Self-supervised vision, Human-in-the-loop retrieval. Links: Project https://realize-lab.github.io/CHAIR · Code https://github.com/realize-lab/CHAIR · Talk https://www.youtube.com/watch?v=Czo5YQU4-WY ### VeLLM: multilingual LLMs for the global majority Microsoft Research India, 2022 – 2023. Status: Prototyping. Research Fellow. Co-led the dynamic prompt/model/embedding selection work that became the COLING 2025 paper. Platform work behind LLM applications piloted in schools across India, improving coverage for non-Latin-script and low-resource languages without additional training. Areas: LLMs, Multilingual evaluation, Prompt & model routing. Links: Project https://www.microsoft.com/en-us/research/project/project-vellm/ · Context https://www.microsoft.com/en-us/research/story/advancing-ai-to-meet-needs-of-the-global-majority/ ## Publications ### RELIANCE: Curating and Evaluating Reproductive Health Information on Social Media KDD 2026 (Datasets and Benchmarks Track). Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining Authors: Vaibhav Balloli, Laura Peyton Ellis, Vishala Mishra, Alice M Chi, Alex Friedman Peahl, Elizabeth Bondi-Kelly. DOI: 10.1145/3770855.3817447. arXiv: 2606.18285. Devise data curation multi-modal agents to collect and curate reproductive health information from TikTok and work with clinicians to evaluate the accuracy of the content and how LLMs perform in evaluating the content for accuracy and harms. Canonical page: https://profile.vballoli.com/papers/reliance/ ### PATHFinder Agent for Tailored Prenatal Care Interactive Health 2026. Proceedings of the 2026 ACM Interactive Health Conference Authors: Vaibhav Balloli, Carissa Samuel, Samia Abdelnabi, Alex Peahl, Elizabeth Bondi-Kelly. DOI: 10.1145/3786579.3804996. arXiv: 2607.24768. Interactive LLM agents that converse with pregnant people to provide personalized prenatal care information, planning, and resources to patients, and to assist clinicians. Canonical page: https://profile.vballoli.com/papers/pathfinder/ ### "Where is this coming from?" Uncovering Trustworthiness Ideals in AI-powered Peripartum Information Seeking FAccT 2026. Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency Authors: Vaibhav Balloli, Julia Erickson, Xinyi Li, Erin MacMurray van Liemt, Alex Friedman Peahl, Elizabeth Bondi-Kelly. DOI: 10.1145/3805689.3812277. arXiv: 2606.10158. Conduct focus groups to understand the role of AI-enabled information provision in the current health ecosystem and the trustworthiness and governance expectations of stakeholders. Canonical page: https://profile.vballoli.com/papers/sociotechnical-design/ ### SEEK-CBM: Editable and Interpretable Retrieval for Elephant Re-Identification CV4Animals @ CVPR 2026. 6th Workshop on CV4Animals: Computer Vision for Animal Behavior Tracking and Modeling, in conjunction with CVPR 2026 Authors: Antoine Salaün, Vaibhav Balloli, Timm Haucke, Elizabeth Bondi-Kelly, Sara Beery. SEEK-CBM upgrades CHAIR to modern vision architectures and self-supervised techniques to achieve a 497% improvement in performance. Canonical page: https://profile.vballoli.com/papers/seek-cbm/ ### Bridging the Gap: Dynamic Learning Strategies for Improving Multilingual Performance in LLMs COLING 2025. Proceedings of the International Conference on Computational Linguistics (COLING) Authors: Somnath Kumar, Vaibhav Balloli, Mercy Ranjit, Tanuja Ganu, Kabir Ahuja, Sunayna Sitaraman, Kalika Bali. arXiv: 2405.18359. Distinctions: Oral. A dynamic strategy that selects, per query at run-time, the optimal prompt, LLM, and embedding model to maximize multilingual QA performance, improving non-Latin and low-resource language coverage without any extra training. Canonical page: https://profile.vballoli.com/papers/leap/ ### Are They the Same Picture? Adapting Concept Bottleneck Models for Human-AI Collaboration in Image Retrieval IJCAI 2024 (Human-Centered AI track). Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 Authors: Vaibhav Balloli, Sara Beery, Elizabeth Bondi-Kelly. DOI: 10.24963/ijcai.2024/866. Distinctions: <5% accept. CHAIR adapts Concept Bottleneck Models so experts can intervene on intermediate concepts during image retrieval, a setting that matters for wildlife re-identification platforms like ElephantBook. Flexible levels of intervention let domain experts spend effort where it's most useful, achieving genuine human-AI complementarity. Canonical page: https://profile.vballoli.com/papers/chair/ ### Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception NeurIPS 2023. Advances in Neural Information Processing Systems (NeurIPS) Authors: Anurag Ghosh, Vaibhav Balloli, Akshay Nambi, Aditya Singh, Tanuja Ganu. DOI: 10.52202/075280-2429. A learned approximate-execution framework for streaming perception. Chanakya jointly considers scene content and system contention to pick run-time decisions (resolution, model, compute) that balance accuracy and latency, beating static and dynamic baselines on both server GPUs and edge devices. Canonical page: https://profile.vballoli.com/papers/chanakya/ ## Topics ### AI for Social Impact Deployments where the measure of success is whether something changed for the people involved: road safety testing across India, elephant re-identification for conservation partners, prenatal care planning with clinicians in the loop. The recurring lesson is that the deployment constraints reshape the model, not the other way round. 4 papers: https://profile.vballoli.com/topics/social-impact/ ### AI for Health Research on AI in peripartum and reproductive health, spanning the full arc from formative study to deployed system to benchmark: what birthing people, clinicians, doulas, and community health workers actually need from an information tool, an agent built to those requirements, and a dataset for checking whether the surrounding information ecosystem is accurate. 3 papers: https://profile.vballoli.com/topics/healthcare/ ### Large Language Models Work on making large language models useful when the stakes are real: choosing prompts, models, and embeddings at run time rather than fixing them in advance, building conversational agents that gather context before they advise, and measuring where frontier models still fall short. The thread running through it is that an LLM is a component in a system, not the system itself. 3 papers: https://profile.vballoli.com/topics/llms/ ### Computer Vision Vision systems designed to be steered by the people who use them. This includes concept-bottleneck retrieval for wildlife re-identification, where a domain expert can intervene on intermediate concepts mid-search, and adaptive streaming perception that trades accuracy against latency at run time based on scene content and system contention. 3 papers: https://profile.vballoli.com/topics/vision/ ### Evaluation & Benchmarks Datasets and evaluation protocols built with the domain experts who have to trust the result. Rather than scoring models on proxies, this work curates expert-annotated ground truth (clinicians reviewing health claims, rubrics written by practising obstetricians) and asks whether language models can stand in for that judgement. Usually the answer is partly, and the interesting part is where the gap sits. 2 papers: https://profile.vballoli.com/topics/evals/ ## News - 2026-04-20: One paper accepted to KDD 2026 in the Datasets and Benchmarks track. - 2026-04-20: One paper accepted to CV4Animals @ CVPR 2026, in collaboration with BeeryLab at MIT. - 2026-04-15: One paper accepted to FAccT 2026. - 2026-04-01: One paper accepted to Interactive Health 2026. - 2025-05-15: Started my internship at AI Frontiers, Microsoft Research. - 2025-05-01: Received "Exceptional" on my Preliminary exam (advancement to candidacy). - 2025-02-15: Attending and presenting a poster at the Society-Centered AI Workshop at Google. - 2025-01-15: Officially a Ph.D. Candidate! - 2024-10-15: One paper accepted at COLING 2025 (Oral): Bridging the Gap: Dynamic Learning Strategies for Improving Multilingual Performance in LLMs. - 2024-05-16: CHAIR accepted at IJCAI 2024, Human-Centered AI Track (<5% acceptance rate). - 2024-04-20: One paper accepted at the CV4Animals Workshop @ CVPR 2024. - 2024-04-10: Second place at the Google × MHacks hackathon, building the next generation of package management in the age of generative AI ($1,500 prize, solo team). - 2023-09-21: Chanakya accepted at NeurIPS 2023. - 2023-08-28: Joined the University of Michigan as a Ph.D. student! - 2023-07-15: Featured on the SARC-BPHC Alumni Unplugged podcast. - 2023-07-10: Selected for the HAIST-MAIA Intro Fellowship on AI Safety. - 2023-07-01: Selected as a Teaching Assistant for the ClimateChange.ai Summer School. - 2023-06-15: Presented research at RF Day, MSR India on decision-making for social good applications. - 2022-07-01: Extending my stint at Microsoft Research as a Research Fellow. - 2021-07-01: Joined Microsoft Research as an SCAI Research Fellow. ## Curriculum vitae ### Research focus - Post-training environments that teach agents to clarify intent, recover, abstain, or delegate. - Contextual evaluations built with domain experts rather than proxy metrics alone. - Understanding and improving agents' instruction-following behavior. - Monitoring deployed systems at test time. ### Research and engineering impact - PATHFinder: Conversational patient- and clinician-facing agent being deployed for personalized prenatal care planning at the University of Michigan. - Elephant Re-ID: Computer vision models being deployed for elephant re-identification in Botswana. - HAMS: Automated driver's-license testing deployed across India; more than 440,000 candidates assessed. ### Technical skills Python, C++, JavaScript; PyTorch, JAX, SkyRL, verl. ### Education - University of Michigan, Ann Arbor (2023 – present (expected Dec 2027)): Ph.D. Candidate, M.S. in Computer Science and Engineering. GPA: 4.0. Advised by Prof. Elizabeth Bondi-Kelly. - BITS Pilani, Hyderabad Campus (2016 – 2020): B.E. in Electronics and Communication Engineering. ### Research experience - Microsoft Research Redmond, AI Frontiers, Research Intern (May – Aug 2025): Applied multi-turn reinforcement learning (GRPO) to help LLM agents clarify intent and recover from errors across extended episodes, and built evaluation harnesses for multi-turn performance. Advised by Hussein Mozannar, Adam Fourney, Gagan Bansal, Saleema Amershi, and Eric Horvitz. - Microsoft Research India, Research Fellow (Jun 2022 – Jun 2023): Built runtime prompt, LLM, and embedding-model selection strategies, plus multilingual evaluation and retrieval components for VeLLM. Evaluated across 18 languages. Also developed Chanakya's contextual-bandit approach to runtime accuracy and latency tradeoffs. Advised by Dr. Akshay Nambi, Tanuja Ganu, and Dr. Venkat Padmanabhan. - Microsoft Research India, Societal Impact through Cloud and AI Fellow (Jun 2021 – Jun 2022): Developed visual SLAM, object detection, and trajectory analysis for HAMS automated driver's-license testing. Advised by Dr. Akshay Nambi, Tanuja Ganu, and Dr. Venkat Padmanabhan. ### Awards and honours - 2025: Exceptional rating, Ph.D. preliminary exam (advancement to candidacy) - 2024: 2nd place, Google × MHacks hackathon (USD 1,500); featured by labs.google - 2023: Selected for the Harvard/MIT HAIST/MAIA introductory fellowship on AI safety - 2022–2024: VeLLM cited by Satya Nadella and covered in the Times of India; HAMS covered in the Punjab News Express - 2023: Featured in a podcast about SmartCampus, co-founded during undergrad ### Service - Conference reviewing: NeurIPS 2024–2026, KDD 2026, IJCAI 2025, ICML 2025, AAMAS 2025, ICLR 2025 - Workshop reviewing: CVPR 2024, NeurIPS 2023, ACL 2026