Ph.D. Candidate · Human AI Synergy Lab · Texas State University
Generative AI should be plausiblegrounded.
I'm a third-year Ph.D. candidate studying how large models fail — and how to catch it. My research detects, measures, and corrects hallucinations in text, vision-language, and audio-language systems: evidence-grounded correction, adversarial stress-testing, and benchmarks that expose silent failure modes before deployment does.
Open to research internships & collaborations
Recent updates
Our project “Trustworthy Vision-Language Models via Minimal Sufficient Captions for Faithful Image Generation” (2026–2027) received 12,400 GPU hours on NCSA Delta through the National Artificial Intelligence Research Resource Pilot (NAIRR260108, PI: Dr. Heena Rathore).
Successfully defended my dissertation proposal.
Cleared the qualifiers in the Department of Computer Science at Texas State University.
Research focus
One question across three modalities: when a model generates something, can we trust it — and if not, can we fix it?
Claim decomposition, verification-question generation, and evidence-grounded rewriting that corrects misinformation without breaking narrative coherence.
Minimal sufficient captions and rate-distortion views of captioning, so downstream image generation stays faithful to what's actually in the scene.
Benchmarking speech-grounded LLMs and tracing how transcription errors propagate quietly into downstream reasoning — especially where the stakes are clinical.
Characterizing prompt-level attacks that induce hallucination, and building evaluations that hold up under distribution shift and adversarial pressure.
Research output
Proceedings of IEEE CCNC, 2026
Hallucination detection designed for LLMs deployed in networked and edge environments, with metrics that account for bandwidth, latency, and resource constraints.
IEEE DISTILL, 2025
An ensemble of classifier, semantic-similarity, NLI, and LLM-judge signals that detects paragraph-level hallucinations and guides targeted revision of long-form outputs.
Proceedings of CIOCSE, 2025
Decomposes paragraphs into verifiable claims, retrieves external evidence, and gates targeted LLM rewrites — correcting misinformation while preserving narrative coherence.
Proceedings of IEEE MobiSecServ, 2024
Quantifies how weak-semantic and out-of-distribution prompt attacks raise hallucination rates in open-source models such as Vicuna-7B and LLaMA-2-7B-chat.
Full list on Google Scholar.
Selected results
Paragraph-level misinformation correction while holding FactCC above 0.97, evaluated across 1,400+ passages.
Weak-semantic prompt attacks reliably induce hallucination in open-source LLMs — the failure modes my detection work targets.
LSTM, custom Transformer, and fine-tuned BERT readers on SQuAD and Natural Questions — over 20 points above TF-IDF baselines.
Joint uplink/downlink resource allocation under high mobility, with a 25% gain in spectral efficiency.
Background
Doctoral Instructional Assistant · Expected 2027
Research Assistant, Wireless Communications Lab
Beyond the lab
I hike, play badminton, and take dance workshops. Lately I've been learning landscape photography — natural light, open skies, quiet places.
Research collaborations, internship conversations, or questions about a paper — my inbox is open.
or find me on LinkedIn and Google Scholar
San Marcos, Texas