I'm an AI researcher at Lossfunk. I work on world models - how an agent builds a picture of its world, and how it keeps learning when the rewards almost never come.
Fine-tuning 0.5B-7B LLMs with QLoRA and 4/8-bit quantization, deployed on Jetson Nano. Active threads: diffusion language models and aggressive compression (PTQ/QAT/AWQ/GPTQ/pruning) feeding Indic text generation.
Mechanistic interpretability with SAEs: attention circuits, feature attribution, and the full SAE gallery from the KD paper. See the gallery.
Study of 18 vision models for PCOS ultrasound. A hybrid CNN-ViT performed best, and a three-stage deduplication pipeline cleaned much of the public dataset. Read more.
Four results on 1 GPU: a 1 µs chignolin fold, titin I27 force-pulling, an 8-protein ESMFold sweep in 12.9 s, and ESMFold calibration. Read more.
Computer vision for detection and autonomous navigation in real-world flight settings.
Teleoperated robotic arm focused on low-latency control and precise manipulation.
World models and continual learning for game-playing agents in sparse-reward environments.
Research on knowledge distillation, representation learning, world models, and interpretability in a student-founded lab. Read more.
Agentic RAG, document intelligence, an internal MCP server, gesture recognition, and supply-chain AI for aerospace. Read more.
Department-led research across language models, autonomous drones, humanoid robotics, and medical AI. See all DSU work.
Built a portfolio of 24 Python projects: data visualisations, Tkinter GUIs, and an end-to-end chatbot, with project management on GitHub and Notion.
Lab founded by students, working on knowledge distillation, world models, and frontier AI. Read more.
Leading tech initiatives for entrepreneurial ventures: overseeing infrastructure and bridging business ideas with technical solutions.
Core Committee, Institution of Engineering and Technology (IET) DSU chapter and the Roboverse student community; ran 10+ tech talks, workshops, and hackathons bridging theory with practice in AI and robotics.
Executive Committee, IEEE RAS and IEEE CIS at DSU. 50+ workshops, talks, and hackathons. RAS & CIS Exceptional Volunteering Award, 2025, given to the top 10% of undergraduate researchers at the university.
NVIDIA Deep Learning Institute course on transformer fundamentals: attention, tokenization, and downstream NLP tasks.
AWS Cloud Practitioner EssentialsFoundational AWS knowledge: core services, security, billing, and the shared-responsibility model.
AWS Technical EssentialsHands-on coverage of compute, storage, databases, networking, and security on AWS.
Microsoft Learn: Benefits of Cloud ArchitectureCloud-architecture fundamentals: design principles, trade-offs, and Azure mappings.
Kaggle: Intro to Machine LearningKaggle's foundational ML course: model fitting, evaluation, and validation on tabular problems.
Altair RapidMiner Certified Data Science ProfessionalProfessional certification covering end-to-end data-science workflows on the Altair RapidMiner platform.
Artificial Intelligence and Robotics Engineering (2023-2027). Designed and taught the LLM Engineering curriculum and lab manuals in my 5th semester, published on the university website - read the manual.
Inside Erdős AI Lab, the day-to-day research stack: knowledge distillation, mechanistic interpretability (SAEs), world models, and protein structure prediction. Includes a KD paper and the full SAE experiment gallery.
AI × Bio: Protein FoldingFour experiments on real proteins, all on 1 GPU: a 1 µs chignolin fold, titin I27 force-pulling, an 8-protein ESMFold sweep, and ESMFold calibration. Reproducible from the linked code.
India AI Impact Summit 2026Representing DSU at India's premier AI conference in New Delhi, engaging with 250+ companies, and presenting research on LLM architecture, medical AI, and autonomous drones.
SAE Experiments: KD paperToy-model validation of the KD minimum-width theorem plus three real-LM trials on Pythia-410M (24 layers × 6 checkpoints, full sparsity sweep) published on Hugging Face.
Language Modeling ExperimentationFine-tuning Qwen 0.5B-7B for low-resource Indic languages; deploying Gemma 3 1B on Jetson Nano. Diffusion-LM and compression threads now live.
Medical AI: PCOS DetectionBenchmark of 18 vision models on PCOS ultrasound. A hybrid CNN-ViT hit 99.81%. A three-stage dedup pipeline cleaned 70% of the public dataset.
If you're working on world models, interpretability, or anything nearby, I'd like to hear about it.