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.