![]() |
MSc Computing (AI and Machine Learning) · Imperial College London |
I'm an MSc student at Imperial, working on large language models and natural language processing. I look at how models represent words and meaning, how one prompt can give several different answers, and how to make models smaller without losing quality.
Before Imperial I studied Computer Science and Engineering at KL University, Hyderabad (CGPA 9.72 / 10). I was a research intern at IIIT Hyderabad, and I also built integration software for enterprise clients at InfoAjax.
I'm looking for funded PhD positions in language models and NLP, starting October 2027. Email me at gk326@ic.ac.uk.
| 2026 – 2027 | MSc Computing (Artificial Intelligence and Machine Learning) Imperial College London |
|
| Oct – Nov 2025 | AI Integration Engineer InfoAjax Consulting · remote I built a proof-of-concept browser agent that acts on screenshots, using models on Azure AI Foundry. I also built a React and FastAPI console to watch agents live. |
|
| Dec 2024 – Jun 2025 | Research Intern IIIT Hyderabad I designed a document translation system for English, Hindi, Telugu and German that keeps the original layout. It uses Vision Transformer layout detection, font detection, colour extraction and Tesseract OCR. I demonstrated it at the institute's research expo (demo). |
|
| Oct 2024 – Jun 2025 | Software Engineer (freelance) InfoAjax Consulting · remote I built REST APIs in WSO2 Micro Integrator that link the service-desk, billing and inventory systems of a Philippine telecoms company. I supported them in production. |
|
| 2024 | Machine learning programme (six months) iHub-Data, IIIT Hyderabad |
|
| 2021 – 2025 | B.Tech Computer Science and Engineering KL University, Hyderabad · CGPA 9.72 / 10 |
- Oct 2026: I tested a redesigned Prism Tuning in pre-registered GPU experiments. It gave no significant gain, and I have a suspected cause.
- Oct 2026: Version 2 of BIE is on Zenodo. It has a new storage format and a full measured evaluation.
- Sep 2026: Started the MSc in Computing (AI and Machine Learning) at Imperial College London.
- Jul 2026: My team won first place at Imperial's Global Challenge Lab, a 14-day sprint with over 1,000 students. I led the AI work on OrbitOps, a reliability layer for satellite onboard AI.
- 2026: Dual-Stream accepted at IEEE NEPCON 2026.
- 2025: MANCE published at IEEE ICCCMLA 2025 (IEEE Xplore).
| Morphology-Aware Nested Character Embeddings for Word Representation Gruhesh Sri Sai Karthik Kurra, Chandanasree Moparthi, Prathipati Manish Chowdary, Pavan Kumar Pagadala IEEE ICCCMLA 2025 · Published in IEEE Xplore A character-level Transformer builds every word from its characters. Related forms share part of their representation, and unseen words still get a vector. In a seeded re-implementation (3 seeds) it beats word embeddings by 14.9 F1 on German NER (63.0 vs 48.2) and is on par on English topic classification. Paper · Code · Results |
|
| Dual-Stream Artifact Detection with Iterative Evidence Refinement for Frame-Level Deepfake Recognition Gruhesh Sri Sai Karthik Kurra, Pavan Kumar Pagadala, Malathy Batumalay, Veda Boddapati IEEE NEPCON 2026 · Accepted, presentation December 2026 A ViT-Small model with an edge stream for blending boundaries and a frequency stream for generator fingerprints, fused by iterative cross-attention. An audit found that 97.3% of WildDeepfake test frames came from videos also used in training. So the main result is 96.3% AUC on a video-disjoint subset of 5,680 frames (87.1% on Celeb-DF), from a single training run. Code · Dataset |
|
| Prism Tuning: Repulsion-Trained Seed Embeddings in a Frozen Transformer for Non-Redundant Generation Gruhesh Sri Sai Karthik Kurra Preprint · 2026 Only K seed vectors are trained. They are added to the input embeddings of a frozen transformer and pushed apart by a repulsion loss, so one prompt gives K different answers in one batched pass, with no filtering. The paper gives the method and formulation only. In October 2026 I tested a redesigned version in pre-registered GPU experiments on Qwen2.5. It did not significantly beat tuned min-p sampling, and the main suspected cause is the training data. |
|
| BIE: Bit-Index Encoding of Sparse Neural Network Weights Gruhesh Sri Sai Karthik Kurra Zenodo preprint · Version 2, October 2026 A storage format for pruned weights. The gaps between non-zeros are Golomb-Rice coded and the values are 8-bit codes stored as bitplanes. At 95% sparsity it is 44.5× smaller than float32 at 0.6% error, against 25.7× for CSR with the same values. Below about 50% sparsity it is no smaller than dense int8. Paper · Code · Results |
The MANCE and BIE repositories include the configs, seeds and per-run result files behind these numbers, so you can re-run each one. More on Google Scholar.
- First place, Global Challenge Lab, Imperial College London, 2026
- First place, University Webathon, KL University, 2022
- Technical Lead, Student Activity Council, KL University, 2023 to 2025: mentored 20+ students and ran workshops for 100+
- Hackathon mentor, Anurag University, Hyderabad
Programming languages: Python, JavaScript, TypeScript, Swift, Java, C, C++, R, SQL, PHP, HTML, CSS, Shell
AI / ML frameworks: PyTorch, TensorFlow, Hugging Face, LangChain, LlamaIndex, scikit-learn, PyTorch Geometric, Gradio, Streamlit, UMAP
AI / ML specialisations: LLM optimisation, SLM fine-tuning, RAG, prompt engineering, Transformers, attention mechanisms, GANs, VAEs, diffusion models, GNNs, on-device AI, deepfake detection, quantisation, pruning, distillation
Computer vision: EfficientNet, Vision Transformer, MediaPipe, OpenCV, Tesseract, CLIP, document analysis, OCR
iOS / mobile: Swift, SwiftUI, Core ML, iOS development, Vision framework, AVFoundation, Core Data, Metal MPS, on-device inference
Web and backend: React.js, Next.js, FastAPI, Node.js, Express.js, TailwindCSS, Vite, React Native, WebSocket, REST APIs
Cloud and DevOps: AWS, Azure OpenAI, Azure AD, Google Cloud, Docker, Vercel, WSO2, CI/CD, Git, GitLab, serverless
Databases and data science: PostgreSQL, MySQL, MongoDB, Redis, Firebase, Supabase, Pandas, NumPy, Matplotlib, FAISS, synthetic data
I write AI from Scratch, a blog series that derives machine learning models from the maths to the training loop. Recent posts:
What We Know About How Jev Works, and How the Open Copies Build ItColibri Explained: How a 744B LLM Runs From Disk
Convolutional Neural Networks: CNN Math Explained
DeepSeek V4 Inside: One Token Through Every Block
Looped Transformers Explained: Recurrent Depth and Astra
























