I build AI systems from scratch.
I’m an incoming MSc Computing (Artificial Intelligence and Machine Learning) student at Imperial College London, starting September 2026.
My research interests include language models, computer vision, neural compression, and on-device AI. In July 2026, my team won first place at Imperial’s Global Challenge Lab with OrbitOps.
InfoAjax Consulting Pvt Ltd · Remote, Hyderabad
Built an integration layer between enterprise apps and cybersecurity services over standardised APIs.
Used Azure OpenAI agents for automated schema mapping and connection setup from prior runs.
Built a React 18 + FastAPI + WebSocket app for live agent monitoring.
Wired Azure AD via Microsoft Graph API for app discovery and metadata.
IIIT Hyderabad · Hyderabad, India
Led Layout Preserving Document Translation: Vision Transformer layout detection (92% accuracy), character-metrics font detection, K-means colour palette extraction (88% accuracy).
Built a FastAPI backend with Tesseract OCR for multi-language text extraction (English, Hindi, Telugu, German) while preserving document formatting.
Presented at the IIIT Hyderabad research expo (May 2025) with live demos to faculty and industry visitors.
InfoAjax Consulting Pvt Ltd · Remote, Hyderabad
Built and shipped RESTful APIs for PLDT (Philippines' largest telecoms provider) ticketing integration in WSO2 Integration Studio.
Handled production secrets via GitLab vault and Choreo cloud hosting.
Debugged production issues with sparse WSO2 docs; worked directly with international clients under strict SLAs.
iHub-Data, IIIT Hyderabad · Hyderabad
Six-month research apprenticeship: neural architecture design, LLM fundamentals, prompt engineering, fine-tuning, quantisation, and deployment.
Document analysis, OCR, and computer vision projects under faculty mentorship.
Zynthetix · Hyderabad, India
Built a synthetic-data platform (GANs/VAEs) for privacy-preserving tabular, image, and text datasets.
Designed a meta-controller where a parent model coordinates 200–1000 specialised child models to generate millions of rows without duplication.
Prototype with transformer-based PII detection (98.5% accuracy) and 95% statistical similarity between synthetic and real data.
Xtraleap India Pvt Ltd · Remote, Hyderabad
Exploratory data analysis, statistical modelling, and ML model development using Python (pandas, NumPy, scikit-learn, Matplotlib).
Feature engineering and model evaluation on real-world datasets in retail analytics and customer segmentation.
I completed my B.Tech in Computer Science and Engineering at KL University, Hyderabad, with a CGPA of 9.72 / 10 (August 2021 – May 2025).
I’m Gruhesh Sri Sai Karthik Kurra, an independent AI/ML researcher, engineer, and founder of Zynthetix. I work on transformer architectures, word representation, neural network compression, and deepfake detection. My work includes an IEEE ICCCMLA 2025 publication, open research on Zenodo, and two papers under review at IEEE NEPCON 2026.
I write about machine learning from scratch on my blog and share my implementations on GitHub.
Boundary and frequency streams under an orthogonality penalty, fused by iterative cross-attention. 96.3% AUC on a video-disjoint split over 227,504 frames.
AUC 96.3% · Frames 227k · Celeb-DF 87.1%
Replicated Attention Is All You Need with BLEU benchmarks, then a GPT-2 (124M) from scratch with BPE and a full training loop.
Paper model Vaswani · GPT-2 124M · Tokeniser BPE
More implementations and experiments. All repositories are on GitHub.
Neural weight compression at 40× with MSE < 10⁻⁶. Zenodo 2025 release with open-source sparse kernels and a benchmarking suite.
Compression 40× · MSE <10⁻⁶ · Year 2025
PubMedBERT pipeline for trial matching. 99.5% accuracy on biomedical similarity with domain embeddings and FAISS retrieval.
Accuracy 99.5% · Encoder PubMedBERT · Index FAISS
GANs trained on MNIST with loss curves, generation metrics, and latent-space interpolation.
Dataset MNIST · Models 2 · Stack PyTorch
DDPM from math to generated images: ~130K-parameter UNet, noise schedule, and sample generation.
Params ~130K · Type DDPM · Net UNet
Privacy-first deepfake detection: retrieval-augmented generation plus visual inconsistency analysis, with an uncertainty-aware classifier.
Year 2024 · Method RAG+CLIP · Status Published
Vision Transformer layout detection plus multi-language OCR (English, Hindi, Telugu, German), keeping fonts, colours, and structure.
Layout acc. 92% · Languages 4 · Lab IIITH
I write about implementing machine learning models from scratch, from the mathematics to the training loop.
Machine learning from scratch · My blog
Posts on dev.to
DeepGuard demo · On-device deepfake detectionLayout-preserving document translation demoAccepted
IEEE ICCCMLA 2025 · 2025
MANCE represents each word as an ordered sequence of shared character embeddings, allowing related word forms to share information without explicit morphological supervision. The paper develops separate RNN, CNN, and hierarchical Transformer integrations and evaluates them on text classification, sentiment analysis, and named entity recognition. Reported results include 93.5% accuracy on AG News, 99.0% on DBpedia, and 81.3% F1 on German named entity recognition. In a separate controlled sentiment experiment with 50 samples, MANCE achieves 91.67% accuracy versus 33.33% for Word2Vec, with training times of 0.51 and 1.11 seconds respectively. These small-experiment gains are distinct from the broader benchmark results.
Accepted
Zenodo · September 2025
Bit-Index Encoding (BIE) compresses sparse neural network weight matrices by recording the positions of nonzero bits. Three variants address different trade-offs: binary encoding for sparse binary weights, bitplane encoding for adjustable reconstruction precision, and blocked encoding for cache locality and parallel processing. The implementation includes Numba-compiled sparse matrix multiplication kernels that operate on compressed representations. Matrix benchmarks cover dimensions from 256 × 256 to 2048 × 1024 and sparsity levels up to 95%. Binary encoding reaches 40× compression at 95% sparsity; separately, bitplane variants achieve reconstruction mean squared error below 10⁻⁶. These results concern matrix compression and reconstruction, rather than downstream model accuracy.
AcceptedZenodo · August 2024
This paper proposes a local deepfake detection framework that combines retrieval-augmented generation with visual inconsistency analysis. A retrieval pipeline uses Vision Transformer embeddings and FAISS to compare an input with known authentic and manipulated samples, then supplies retrieved evidence to a local language model. A complementary visual pipeline uses object detection and segmentation to examine anatomical irregularities, facial symmetry, boundaries, and lighting. Confidence-weighted fusion combines the two sources of evidence, while uncertainty estimates identify ambiguous cases for human review. The paper specifies the architecture and evaluation protocol for privacy-preserving analysis. Its performance figures are projections; implementation and full empirical validation remain future work.
Under ReviewIEEE NEPCON 2026 · Submitted 2026
Prism Tuning proposes a parameter-efficient method for generating diverse responses from a frozen language model. K learnable seed vectors are added to input embeddings to condition parallel generation streams. During training, a repulsion objective acts on generated-position hidden states, providing a differentiable proxy for differences between decoded outputs. Whitening addresses representation anisotropy, while shared style parameters and separate content parameters aim to discourage diversity based only on tone. Quality and KL-divergence terms anchor the streams to the base model. At inference, the seeds require no coordination between streams, though generating K responses still incurs their generation cost. The paper presents the formulation and integration requirements; systematic empirical evaluation remains future work.
Under ReviewIEEE NEPCON 2026 · Submitted 2026
This study combines a Sobel-edge boundary stream with a multi-band Fourier frequency stream for frame-level deepfake detection. An orthogonality penalty discourages redundant evidence, and an iterative cross-attention module refines and fuses the streams. The study uses a corpus of 227,504 frames from FaceForensics++, Celeb-DF v2, and WildDeepfake. A source-video audit found extensive train–test overlap in WildDeepfake, so the paper separates the contaminated evaluation from a video-disjoint subset. On that 5,680-frame subset, the model achieves 96.3% AUC and 90.5% accuracy; Celeb-DF alone yields 87.1% AUC. Results come from one training run, without controlled baseline comparisons or ablations, which limits claims about the contribution of individual components.
Code and dataset: GitHub · Kaggle dataset.
RADAR proposes deepfake detection through the aggregation of complementary forensic evidence. Three specialised branches examine skin texture, boundary aliasing, and fingerprints associated with AI image generation. An R-Former module combines their outputs through iterative cross-attention, with per-artifact confidence scores and spatial saliency maps intended to explain predictions. The design specifies a Vision Transformer trained from random initialisation, alongside artifact supervision and an evaluation protocol covering detection, cross-dataset generalisation, and interpretability. A 200-image pilot analysis motivates the artifact branches. Full-system benchmark validation remains future work, and the paper’s performance figures are design targets rather than measured outcomes. The proposed scope is facial manipulation in individual images.
Won Imperial's global innovation sprint: 1,000+ students, 14 days, four tracks. Built OrbitOps with a five-person team, a vendor-neutral reliability layer for satellite onboard AI (fault injection before launch, known-answer probes in orbit, rollback on silent degradation). £1,500 team prize; certificate in Certifications.
Diet management platform in a 4-hour hackathon: personalised meal planning, calorie tracking, and nutritional recommendations. Competed against 50+ teams.
Arduino smart switchboard for IoT home automation with real-time energy monitoring and mobile app control.
Technical Lead: mentored 20+ juniors in web development and ML, ran workshops for 100+ students, coordinated 15+ technical events.
Google · 2024
Amazon Web Services · 2023
Amazon Web Services · 2023
Oracle · 2023
Oracle · 2024
Oracle · 2023
Red Hat · 2024
WSO2 · 2025
Hugging Face · 2024
Hugging Face · 2024
IIIT Hyderabad · 2024
WSO2 · 2025
Google Developer Groups · 2025
Udemy · 2026
Udemy · 2026
IELTS Academic: overall band 7.5 · C1 Advanced
Python
JavaScript
TypeScript
Swift
Java
C
C++
R
SQL
PHP
HTML
CSS
Shell
PyTorch
TensorFlow
Hugging Face
LangChain
LlamaIndex
scikit-learn
PyTorch Geometric
Gradio
Streamlit
UMAP
LLM Optimisation
SLM Fine-tuning
RAG
Prompt Engineering
Transformers
Attention Mechanisms
GANs
VAEs
Diffusion Models
GNNs
On-Device AI
Deepfake Detection
Quantisation
Pruning
Distillation
EfficientNet
Vision Transformer
MediaPipe
OpenCV
Tesseract
CLIP
Document Analysis
OCR
Swift
SwiftUI
Core ML
iOS Development
Vision Framework
AVFoundation
Core Data
Metal MPS
On-Device Inference
React.js
Next.js
FastAPI
Node.js
Express.js
TailwindCSS
Vite
React Native
WebSocket
REST APIs
AWS
Azure OpenAI
Azure AD
Google Cloud
Docker
Vercel
WSO2
CI/CD
Git
GitLab
Serverless
PostgreSQL
MySQL
MongoDB
Redis
Firebase
Supabase
Pandas
NumPy
Matplotlib
FAISS
Synthetic Data
I’m open to research collaborations, consulting, and full-time roles in AI engineering and ML systems. Hyderabad or remote.
gruheshkurra2@gmail.com · LinkedIn · Download CV
Transformer research · LLM development · On-device AI · Vision-language models · Neural compression · Agentic systems