Bhanu Prakash Vangala
Ph.D. Candidate, Executable Reliability for AI Systems
When trust, reproducibility, and efficiency are the same problem
Electrical Engineering & Computer Science · University of Missouri · Columbia, Missouri, USA
Advised by Dr. Tanu Malik. Ph.D. expected June 2027.
NASADoDNSFGoogle PhD Nominee '25Microsoft Research Intern '26AAAI '25, '26Outstanding MS Award '25
I work where machine learning meets the messy reality of the systems that have to run it. One question keeps pulling me back. When an AI pipeline hands you a claim or an artifact, what does it take to actually trust it? Models today are fluent and confident and often wrong, in ways that are genuinely hard to catch, and the code they write tends to run on the machine that produced it and nowhere else. My dissertation, executable reliability for AI systems, makes a fairly plain argument: whatever a system says about itself, about its outputs, its dependencies or its resource use, someone outside the team that built it ought to be able to check.
Research interests: Agentic AI systems · Multi-model LLM orchestration · Trustworthy & reproducible ML · Distributed systems for inference · Provenance
- 11Selected publications
- 5Oral presentations
- 3Federal funders
- 115+Students taught
The argument
Three threads that look unrelated from the outside. They are one question asked three ways. Can anyone else check this?
-
Trustworthy outputs
Knowing when a scientific answer is invented
A benchmark and a multi-stage detector for hallucination in materials-science question answering, combining self-consistency checking, retrieval against curated sources, and a knowledge graph that flags low-connectivity fact fragments as likely fabrications. It cuts hallucination rates by about 30% and reaches F1 0.95 on cross-domain inconsistency detection. The number is the least interesting part. What I wanted was a pipeline you can interrogate. Ask why a response got flagged and you get a reason back, not a score.
-
Reproducible artifacts
Can anyone else actually run it?
100 standardised prompts given to Claude, Codex and Gemini agents, and the 300 resulting projects executed in clean containers. Only 68.3% ran out of the box: Python 89%, Java 44%. The agents declared about three dependencies per project; the projects pulled in around 49 at runtime. That 13.5x gap does not disappear. It lands on whoever picks the code up next. "It ran on my machine" and "it is reproducible" are two different claims, and our benchmarks have quietly been scoring the first one.
-
Accountable systems at scale
Nine models, a fixed number of GPUs
A Kubernetes-native multi-tenant serving framework that classifies queries by complexity, routes them across nine models from 1B to 27B parameters using Thompson Sampling, and manages warm pools through a custom cold-start state machine. Across 310,190 inference requests on the NRP Nautilus cluster it reached 61% GPU utilization against 31% for static deployment, cutting GPU-hours by 46%. Most of that comes from one decision: stop holding nine models resident when the traffic does not call for them.
News
- 2026.08Completed my Research Data Science internship at Microsoft in Redmond, WA: temporal featurization for predictive Windows retention across 1B+ users.
- 2026.06Pick-and-Spin: Cold-Start-Aware Routing for Self-Hosted LLM Serving was accepted as an oral at IEEE CLOUD 2026.
- 2026.05Beyond Functional Correctness was accepted as an oral at ACM REP 2026.
- 2026.05Submitted CAMP: Consumption-Aware Memory Prediction for Scientific Workflows to IEEE Cluster 2026.
- 2026.04Received the EECS Graduate Travel Fellowship from the University of Missouri College of Engineering.
- 2026.04Awarded a Chameleon Cloud Travel Award (top 10 proposals) for the Sixth Chameleon User Meeting.
- 2026.02My presentation on Evaluating Dependency Gaps in LLM-Generated Code was selected for the Sixth Chameleon User Meeting at NCAR’s Mesa Lab in Boulder, Colorado.
- 2026.01Presented AI-Generated Code Is Not Reproducible (Yet) (Jan 26) and Efficient Multi-Model Orchestration for Self-Hosted LLMs (Jan 27) as orals at the AAAI 2026 RAI and DAI workshops in Singapore.
- 2026.01Submitted Code That Works, Environments That Don’t (an extended 1,000-instance study) to AI Magazine.
- 2025.12AI-Generated Code Is Not Reproducible (Yet) was accepted at the RAI 2025 workshop.
- 2025.11Efficient Multi-Model Orchestration for Self-Hosted LLMs was accepted at the Deployable AI (DAI 2025) workshop.
- 2025.11Received an offer from Microsoft for a Research Data Science Internship (Summer 2026).
- 2025.09Named an Outstanding Reviewer at NeurIPS 2025 (AI for Accelerated Materials Design track).
- 2025.05Graduated with my M.S. in Computer Science and continued into the Ph.D. program at Mizzou.
- 2025.04Selected as a Google Ph.D. Fellowship Nominee (NLP), top 3 of 6,000 participants.
- 2025.04Presented our hallucination-detection work at the AAAI Spring Symposium 2025 (AI for Scientific Discovery track).
- 2025.03Received the Outstanding Master’s Student Award at the Mizzou Engineering Awards Banquet.
- 2025.03Runner-up at the MUIDSI x IBM AI for Social Good Hackathon with VisionAI ($1,000).
- 2025.01Two papers accepted at the AAAI 2025 Spring Symposium: HalluMat and HalluFormer.
- 2024.09Began documenting scalable LLM-as-a-Service infrastructure using Helm charts and node-affinity scheduling on NRP Nautilus.
- 2023.12Joined the Radiant Lab and the Data Intensive Computing Lab as a graduate research assistant.
- 2023.08Began research on faithfulness, interpretability and robustness in large generative language models.
- 2023.06Admitted to the Ph.D. program in Computer Science at the University of Missouri.
- 2023.05Graduated with a B.Tech in CSE (Data Analytics) from VIT Vellore.
- 2023.04Honored with the Dean’s Research Excellence Award at VIT for multilingual NLP and social media analytics.
- 2022.05Started as a volunteer research intern at Adobe Research.
Experience
-
May 2026 – Aug 2026
Research Data Science Intern
Microsoft · Redmond, WA
With Juan-Arturo Herrera, Anqi Cheng, Kyle Hatridge
- Built a reusable temporal featurization framework (TFT / TCN) for predictive Windows retention across 1B+ users, 5,000+ features and 15+ years of Windows and Devices telemetry, a +45% PR-AUC lift over baseline, driven by lag-based trajectory signals with SHAP interpretability.
- Built the Cosmos / SCOPE data layer over Azure streams and data lakes, and automated Copilot Studio feature-analysis workflows.
TFT/TCNAzureCosmos/SCOPESHAPCopilot Studio
-
Dec 2023 – Present
Graduate Research & Teaching Assistant
Radiant Lab & Data Intensive Computing Lab, University of Missouri · Columbia, MO
- Radiant Lab (NASA / NSF). Drove NASA-funded research on LLMs, coding-agent reproducibility and dependency resolution; designed the SciUnit ptrace audit harness; first-authored multiple peer-reviewed papers.
- Data Intensive Computing Lab (DoD / NSF). Built hallucination-detection pipelines; delivered a 40% geospatial-accuracy lift on UAV workflows.
- Kubernetes / Helm on NRP Nautilus HPC. GPU quota management and deployment automation for 9 self-hosted LLMs.
- TA, Web Development (MERN). Mentored 115+ students; built CI/CD scaffolding and grading automation on GitLab.
KubernetesHelmvLLMSLURMptrace
-
May 2022 – Jan 2023
Volunteer Research Intern
Adobe Research · Remote
- Built large-scale web scraping and information-extraction pipelines with automated quality control for ML workflows.
Publications
Google Scholar →Mine is the name in bold. Filter by topic if you want a narrower slice.
-
Code That Works, Environments That Don't: Measuring Environment Reproducibility in AI-Generated Software
Three agents, four languages, fifty tasks. Dependency-set agreement falls as low as 7% on identical tasks, and newer agents show no improvement.
What it's about
Functional correctness says nothing about whether the environment a coding agent declares is the environment its code actually needs. A three-layer framework separates declared, runtime-installed, and necessary-and-sufficient dependencies, and shows the failure is systematic rather than incidental: the widest divergence sits between what an agent declares and what the code pulls in at runtime, which points at environment priors learned during training as the cause.
-
Consumption-Aware Memory Prediction for Scientific Workflows
CAMP: selective eBPF / strace auditing with NGBoost LogNormal predictors; median APE reduction across 35K+ Slurm tasks on AWS ParallelCluster.
What it's about
Scientific jobs are routinely over-provisioned because nobody knows how much memory they will actually need. CAMP audits real consumption cheaply and predicts a distribution rather than a point estimate, so schedulers can reason about risk instead of padding.
-
Beyond Functional Correctness: Evaluating AI-Generated Software Services
A reproducibility-aware evaluation that scores environment specification, build reliability and runtime behaviour, not just whether the tests pass.
What it's about
"The tests pass" is a weak acceptance criterion for a generated service. This paper proposes what else has to hold before you can claim the artifact works.
-
Resource-Aware Multi-Model Serving for GPU Cloud Infrastructure (Pick-and-Spin)
Nine models from 1B to 27B routed by Thompson Sampling with a cold-start state machine. 310,190 requests on NRP Nautilus: 61% GPU utilization against 31% for static deployment, cutting GPU-hours by 46%.
What it's about
Serving many fine-tuned models on a fixed GPU pool forces a choice between keeping everything warm (expensive) and cold-starting on demand (slow). Pick-and-Spin treats routing as a bandit problem and makes residency an explicit, learned decision rather than a static policy.
-
AI-Generated Code Is Not Reproducible (Yet): Dependency Gaps in Coding Agents
100 prompts given to Claude, Codex and Gemini. Of the 300 resulting projects only 68.3% ran in a clean container: Python 89%, Java 44%. Agents declared about 3 dependencies per project; the projects pulled in around 49 at runtime, a 13.5x gap.
What it's about
Code from coding agents runs where it was written and breaks everywhere else. We audited 300 projects generated by Claude, Gemini and Codex and measured how often the declared environment fails to match what the code actually needs.
-
Efficient Multi-Model Orchestration for Self-Hosted LLMs
Multi-agent routing plus GPU orchestration for cost-efficient self-hosted LLM pipelines.
What it's about
How to run a pipeline of specialised models on your own hardware without paying to keep all of them resident.
-
Adaptive Inference: Orchestrating Fine-Tuned LLMs with Serverless GPUs in HPC
Pick-and-Spin router with serverless GPUs on Kubernetes; balances latency against cost and scales to zero.
What it's about
Treats GPU capacity as elastic, so fine-tuned models spin up per request instead of holding hardware indefinitely.
-
HalluMat: Detecting Hallucinations in LLM-Generated Materials Science Content Through Multi-Stage Verification
Self-consistency checking, retrieval against curated sources, and a knowledge graph that flags low-connectivity fact fragments. Cuts hallucination rates by about 30% and reaches F1 0.95 on cross-domain inconsistency detection.
What it's about
A benchmark and detector for hallucination in materials-science LLMs, combining retrieval augmentation (IR, NER and semantic similarity) with intrinsic consistency checks. Introduces the Paraphrased Hallucination Consistency Score, which asks whether a model contradicts itself when the same question is worded differently.
-
HalluFormer: Detecting Hallucination in LLMs
A BigBird span-level classifier: F1 0.947 on MultiNLI and 0.728 on ANAH.
What it's about
Scores consistency across the question, the answer and the retrieved context, rather than judging the answer in isolation.
-
Deploying LLMs as a Service in a Kubernetes HPC Cluster
What it's about
A Helm-based, GPU-aware deployment path for multi-tenant LLM inference on a research Kubernetes cluster, with autoscaling, resource throttling and node-affinity scheduling.
-
Multilingual Sentiment Analysis on KOO User Posts
What it's about
A sentiment pipeline covering ten-plus Indian languages, built for streaming content moderation. Advised by Dr. P. Kumaraguru (IIIT Hyderabad). Won the Best Department Thesis award.
-
AI-Powered Image Colorization: From Grayscale to Photorealistic Colour
What it's about
Convolutional autoencoder and GAN-based colorization with a combined perceptual and adversarial objective, applied to photograph restoration.
-
Scores That Hold, Benchmarks That Leak: Measuring Dataset Contamination in Public Brain-Tumor MRI Classification
28.8% of the dominant corpus's official test split has a near-twin in its own training split; a second corpus leaks 22.3% byte-identically; 95.5% of traceable test images share a patient with training; file headers containing no anatomy separate tumour from no-tumour at 0.959 balanced accuracy.
What it's about
Brain-tumour MRI classification routinely reports accuracies above 98% on a handful of public corpora. A three-layer contamination framework covering duplicate, patient and source-label leakage audits those corpora against a chest-radiograph negative control, across nine architectures. The uncomfortable finding is that removing every leaked test image leaves balanced accuracy essentially unchanged: contamination this severe does not show up as an inflated score, so deduplication repairs the number without repairing the benchmark.
-
Evaluation Choices Shape Biomedical ML Claims: A Pediatric Pneumonia Benchmark Case Study
Eight pretrained backbones differ by only 0.026 AUROC, but freezing versus fine-tuning moves AUROC by 0.044 and changing the decision threshold moves balanced accuracy by 0.090. A file-properties-only classifier with no anatomy reaches 0.992 in the training pool and 0.496 on the official split.
What it's about
A biomedical paper is often remembered by one number, and that number reads as a property of the model when it is really a property of the evaluation. On the Kermany pediatric chest radiograph benchmark, the choice of split, training policy, threshold, metric and calibration moves the headline more than the choice of architecture does. Ends with a seven-item reporting recommendation, each item tied to an effect measured in the study.
No publications match that filter.
Research programs
The funded programs this work lives inside, and the people I do it with.
-
Reproducible Containers for Collaborative Analytics
With Dr. Tanu Malik (University of Missouri), Dr. Ashish Gehani (SRI International), Dr. Kwo-Sen Kuo (NASA), Dr. Abhilash Jindal (IIT Delhi)
- FlexiFlow
- Bandit-based model switching, FastAPI routing with GPU telemetry, Prometheus + Grafana observability. Demonstrated at SIGMOD.
- Pick-and-Spin
- Thompson Sampling with a Cold-Start State Machine: 9.2% cold-starts across 310K runs and 61% GPU utilization, against 31% for the static baseline.
- Self-hosting LLMs
- Helm + vLLM LLM-as-a-Service on NSF Nautilus with KEDA autoscaling; serverless on AWS SageMaker and Lambda with scale-to-zero.
- Coding-agent reproducibility
- 300 repositories generated by Claude, Gemini and Codex, replayed via SLURM; the SciUnit ptrace harness extended to Java and JavaScript.
-
TRACE: Autonomous Experimentation Lab using Agentic AI
With Dr. Tanu Malik (University of Missouri), Dr. Matt Maschmann (University of Missouri), Dr. Derek T. Anderson (University of Missouri)
- MCP & tooling
- Authored custom MCP servers and agent skills wrapping Nanoscribe and SEM instrument APIs; ran 50+ automated experiment cycles with per-call provenance.
- Multi-agent orchestration
- Coordinated agents with Google ADK and Agent2Agent (A2A) over gRPC and Protocol Buffers, behind a FastAPI control plane with pytest and GitHub Actions CI for reproducible runs.
-
Accelerating Materials Design through AI and Machine Learning
With Dr. Jianlin Cheng (University of Missouri), Dr. Grant Scott (University of Missouri), Dr. Prasad Calyam (University of Missouri), Dr. Chitta Baral (Arizona State University), Dr. Lenore Dai (Arizona State University)
- HalluFormer
- A BigBird span-level hallucination classifier: F1 0.947 on MultiNLI and 0.728 on ANAH.
- HalluMat
- The HalluMatData benchmark with retrieval augmentation over IR, NER and semantic similarity: a 30% hallucination reduction.
Projects
-
LearnLLM.dev
A GeeksforGeeks-style learning platform for building with large language models, built full-stack from schema to deployment, with Vercel CI/CD. 1,000+ active users.
-
VisionAI: AI assistance for visually impaired users
Fine-tuned multimodal LLMs for hazard detection and accessibility from a live camera feed. Runner-up at the IBM / MUIDSI Generative AI for Social Good Hackathon 2025.
-
ReflectMemory: persistent memory for long-context reasoning
Stores and re-injects chain-of-thought embeddings so a long-running agent keeps a stable line of reasoning across sessions instead of restarting it.
-
Autonomous indoor navigation
BLE beacons and a Raspberry Pi running A* and Dijkstra pathfinding, paired with a Kotlin app giving turn-by-turn guidance indoors where GPS fails.
-
ChatMed: grounded medical question answering
A symptom-guidance assistant built on BioGPT and grounded in PubMed articles.
-
CropInsight: crop health and yield forecasting
Computer vision plus sequence models for crop health monitoring and yield forecasting from field imagery.
-
SocialSift: crisis-aware multilingual sentiment analysis
Transformer-based sentiment and intent analysis over multilingual social media during natural disasters.
Honors & awards
-
2025
Google Ph.D. Fellowship Nominee (NLP track)
Top 3 of 6,000 participants
-
2026
EECS Graduate Travel Fellowship
University of Missouri
-
2026
Chameleon Cloud Travel Award, top 10 proposals
NSF Chameleon Cloud Infrastructure
-
2025
Outstanding Reviewer Award
NeurIPS 2025, AI for Accelerated Materials Design (AI4MAT) track
-
2025
Outstanding Master's Student Award
University of Missouri
-
2025
Runner-Up, AI for Social Good Hackathon
MUIDSI x IBM · VisionAI, $1,000
-
2022-2023
Dean's Research Excellence Award, Best Thesis Award, Top-2 Academic Performer
Vellore Institute of Technology
Education
-
Aug 2023 – Jun 2027 (expected)
Ph.D., Computer Science
University of Missouri · Columbia, MO · GPA 3.9 / 4.0
- Thesis area: trustworthy and reproducible LLM-based agentic systems; multi-model orchestration; provenance tracing.
- Advised by Dr. Tanu Malik (Radiant Lab).
- Google Ph.D. Fellowship Nominee (NLP track, 2025); EECS Travel Fellowship ‘26; Chameleon top-10 presentation ‘26.
-
Aug 2023 – May 2025
M.S., Computer Science
University of Missouri · Columbia, MO · GPA 4.0 / 4.0
- Thesis: Deploying LLMs as a Service in a Kubernetes HPC Cluster.
- Advised by Dr. Grant J. Scott and Dr. Jianlin Cheng.
- Outstanding Master’s Student Award (2025); Runner-up, MUIDSI Hackathon (2025).
-
May 2019 – Apr 2023
B.Tech, CSE (Data Analytics)
Vellore Institute of Technology · Vellore, India · GPA 8.8 / 10
- Thesis: Multilingual Sentiment Analysis on KOO User Posts.
- Advisors: Dr. P. Kumaraguru (IIIT Hyderabad) and Dr. Soughbhagya.
- Dean’s Research Excellence Award (2023); Top-2 Academic Performer (2022); Merit Scholarship.
Talks, service & teaching
Invited talks & presentations
- Apr 2026
- Evaluating Dependency Gaps in LLM-Generated Code · Sixth Chameleon User Meeting, NCAR Mesa Lab, Boulder, CO
- Jan 2026
- AI-Generated Code Is Not Reproducible (Yet) · AAAI 2026 RAI Workshop, Singapore
- Jan 2026
- Efficient Multi-Model Orchestration for Self-Hosted LLMs · AAAI 2026 DAI Workshop, Singapore
- Apr 2025
- Hallucination detection in scientific LLMs · AAAI Spring Symposium 2025 · AI for Scientific Discovery
Presenting hallucination-detection work at the AAAI Spring Symposium 2025.
Reviewing
Outstanding Reviewer, NeurIPS 2025 (AI4MAT)ACM CAIS 2026 (Main + AEC)NeurIPS 2026 (Main + AEC)ICLR 2025CIKM 2025IEEE 2024
Teaching
- Fall 2025, Fall 2024, Spring 2024, Fall 2023
- Teaching Assistant, Web Development (MERN stack) · University of Missouri. Mentored 115+ students; built CI/CD scaffolding and grading automation on GitLab.
Get in touch
Happy to talk about agentic systems, reproducibility, or serving LLMs at scale. I'm open to collaborations and to speaking invitations.
Or ask the assistant in the corner. It only knows what's on this page, and it will tell you which section it took each answer from.