🏆 Best AI for Social Impact Tech Spark 2024
MS Applied AI · Stevens Institute of Technology · Dec 2026
Building SLM ensembles where open-source models Mistral, Gemma, DeepSeek vote on domain-specific reasoning, fine-tuned with LoRA on epidemiology datasets and benchmarked on NVIDIA L40s GPUs.
I'm Ram Kasuru an MS Applied AI student at Stevens Institute of Technology, graduating December 2026. My research focuses on parameter-efficient fine-tuning, SLM ensemble architectures, and efficient GPU-based benchmarking on real-world domain data.
I build systems where multiple open-source small language models collaboratively reason and vote on outputs, combining the efficiency of SLMs with ensemble-level robustness. I'm also actively exploring PhD opportunities at the intersection of efficient NLP and applied machine learning.
2+ published · 1 in progress
An end-to-end multimodal pipeline fusing YOLOv8 for real-time object detection, OpenAI Whisper for automatic speech recognition, and GPT-4 for contextual language insight generation — achieving 90% object detection and 98% ASR accuracy on live video streams.
Tensor-based MRI stroke diagnosis using Tucker decomposition combined with EfficientNetB0 transfer learning. Achieves 98% classification accuracy across four pathological tumor categories on a multi-class dataset of 3,000+ MRI scans.
A systematic study building an ensemble of LoRA fine-tuned open-source Small Language Models — Mistral 3B, Gemma, and DeepSeek — each trained on a domain-specific epidemiology tutoring dataset. Models cast votes on final outputs, aggregating domain-specific reasoning across architectures. Demonstrates 30%+ improvement in factual recall while reducing GPU memory overhead by 60%+ compared to monolithic LLM approaches.
Real-time live video analysis fusing YOLOv8, Whisper ASR, and GPT-4 into a unified accessibility and insight platform. Presented at ISCMCTR-2024.
Multi-model ensemble (Mistral 3B, Gemma, DeepSeek) fine-tuned with LoRA on a synthetic epidemiology dataset. Voting-based aggregation on NVIDIA L40s via SLURM.
CNN + Tucker tensor factorization pipeline for MRI-based brain tumor classification. 98% accuracy with EfficientNetB0 and interactive Jupyter widget inference.
At Outlier AI improved LLM evaluation accuracy by 12% across 10k+ prompts via reinforcement-based tuning; reduced evaluation latency by 18% across 5+ LLM families.
Scalable ML-driven API automation for 200+ clients at Ziberr Communications. Improved user engagement by 40% and operational pipeline efficiency by 28%.
What keeps me curious outside the GPU cluster.
3× Best Delegate at international MUN conferences. Competed and chaired across simulations spanning climate policy, cybersecurity governance, and global health honing structured argumentation and diplomacy under pressure.
Heavy reader across AI research papers, cognitive science, and philosophy of mind. Regular on arXiv particularly tracking efficient inference, mechanistic interpretability, and emergent reasoning in LLMs.
Just starting out in Muay Thai drawn to its blend of technical precision and full-body conditioning. Treating it the same way I approach a new model architecture: fundamentals first, iterate from there.
Competitive FPS COD Warzone, CS:GO, and Valorant for fast reflexes and team coordination. Also learning Chess every day; finding that studying openings and endgames maps surprisingly well onto search-and-planning in AI.
Lived and worked across South Asia before moving to New Jersey. Passionate about cross-cultural exchange shaped by years of international MUN and NGO operations across the region.
Lo-fi, ambient, and film scores as the backdrop for deep work. Believe the right soundscape is a genuine productivity tool treat it with the same intentionality as a well-crafted prompt.
Lakshaya NGO for Model United Nations
South Asia Region
S.py Graduate AI Club
Stevens Institute of Technology
His knowledge spans diverse domains including Recommendation Systems, Generative Adversarial Networks, and Computer Vision, showcasing expertise in Deep Learning. His mastery of technical subjects is complemented by his ability to cultivate critical thinking and problem-solving.
In the VIOLA major project, he demonstrated exceptional research skills and effectively applied acquired knowledge to a cutting-edge, next-generation project. Frequent experimentation with new ideas underscores his dedication to pushing the boundaries of knowledge.
His unique blend of technical expertise, continuous learning, and outstanding soft skills make him invaluable. His passion for leveraging data analytics and AI for real-world applications positions him as a valuable asset with the potential to drive innovation in Data Science.
Open to AI/ML engineering roles, research collaborations, and PhD opportunities. If you're working on efficient language models, multimodal systems, or applied NLP let's talk.
Based in Jersey City, NJ · Available for on-site, hybrid, or remote positions · F-1 OPT eligible upon graduation (Dec 2026)