Hi, I'm
Vaishak Girish Kumar
MS in AI at University at Buffalo. Building hallucination-resistant medical VLMs, LLM observability tools, and neuroscience-inspired architectures.
01 / About
About Me
I'm an AI/ML researcher focused on building systems that are both powerful and trustworthy. My current work centers on hallucination-resistant vision-language models for medical imaging and developer tools for understanding RAG pipeline behavior.
I believe the next frontier in AI isn't just better models. It's models we can actually trust and understand. That's why I focus on grounding, observability, and explainability in everything I build.
Previously contributed to open-source NLP infrastructure at Scitonic and published two peer-reviewed papers while completing my B.Tech at Presidency University.
M.S. Engineering Science (AI)
University at Buffalo · Expected May 2027
B.Tech Computer Science
Presidency University · 2024
Buffalo, NY
Available for remote collaboration
2
Publications
3+
Research Projects
17
OSS Contributions
30%
Perf Improvements
02 / Research
Research & Publications
Focused on making AI systems more reliable and trustworthy, particularly in medical imaging where precision matters most.
FOVEA-Net
In ProgressTargeting MICCAI 2026 Workshop
Hallucination-resistant vision-language model for chest CT. A 5-stage pipeline grounds each clinical finding to 3D spatial evidence, making hallucination architecturally difficult.
- 5-stage pipeline: MedNeXt-L encoder, 3D RPN foveation, local region transformer, confidence-gated decoder, report aggregation
- Target under 15% hallucination rate from a 35% baseline, with mean IoU above 0.70
- Grounded in a 26-paper literature survey
SalienceFormer
CompletedIndependent Research
Neuroscience-inspired transformer that adds hippocampal salience gating and memory consolidation layers to a standard architecture.
- 11.83 perplexity on WikiText-2, 35% better than the Gemma-2B baseline
- 15 ablation variants with statistical significance testing
- Trained weights published to the HuggingFace Hub
Peer-Reviewed Publications
AI-DxMH: Artificial Intelligence Diagnosis for Modern Health
First AuthorV. G. Kumar, M. F. Pasha, A. Prusty, D. Rajeev, G. Ganesan
Peer-Reviewed Publication · 2024
A comprehensive framework for AI-assisted medical diagnosis in resource-constrained environments, achieving 92% accuracy across multiple conditions.
- Developed multi-modal diagnostic pipeline combining patient history, symptoms, and lab results
- Achieved 92% diagnostic accuracy across 15+ common conditions in clinical validation
- Optimized for deployment in resource-constrained healthcare settings with 30% model compression via LoRA
- Implemented explainable AI features for physician trust and regulatory compliance
On-the-fly Prompt Optimization in Multi-Agent Systems: A Comparative Study
Second AuthorM. F. Pasha, V. G. Kumar, A. Prusty, S. Taj
Peer-Reviewed Publication · 2024
Evaluating dynamic prompt optimization strategies in multi-agent LLM architectures for improved task performance.
- Comparative analysis of 5 prompt optimization strategies across multi-agent workflows
- Demonstrated 23% improvement in task completion rates with dynamic prompt refinement
- Introduced feedback-loop mechanism for real-time prompt adaptation between agents
- Benchmarked on complex reasoning tasks requiring multi-step agent coordination
03 / Projects
Featured Projects
A selection of research projects and tools focused on making AI systems more reliable, observable, and trustworthy.
Other Projects
ResearchGraph
Autonomous research operating system that coordinates 13 agents over a topological dependency graph to survey literature, analyze gaps, and draft proposals.
GroundAgent
Agentic chest X-ray analysis on AMD MI300X using Qwen2-VL-7B with a DeBERTa NLI hallucination gate for grounded report generation.
Peptide Diffusion
Mass-constrained diffusion for de novo peptide sequencing, with entropy-adaptive mass gating and spectral noise augmentation.
AI-DxMH
Fine-tuned LLM diagnostic framework for low-resource healthcare settings. Reaches 92% diagnostic accuracy with 30% model compression via LoRA.
04 / Experience
Experience & Skills
Open-Source Developer
Scitonic via Tonic.AI · Remote
Core contributor to LLM pipeline infrastructure through Tonic.AI's open-source program, focusing on NLP tooling and data processing optimization.
Technical Skills
Languages
ML / Deep Learning
LLM / NLP
Infrastructure
05 / Contact
Get in Touch
Interested in collaboration, research opportunities, or discussing AI? I'm always open to connecting with fellow researchers and engineers.