Sohan Reddy
AI & Data Science undergraduate specializing in Deep Learning, NLP, and Retrieval-Augmented Generation (RAG) systems. Experienced in end-to-end model deployment using PyTorch, ONNX Runtime, and quantized LLMs with production implementations across on-device inference, semantic vector search, and real-time voice pipelines. First Runner-up, Bengaluru AI Hackathon
Experience
AI Engineer Intern
CoComply AI
Bengaluru, Karnataka
Developing backend infrastructure and AI capabilities for Agent Coco, a multi-tier compliance AI assistant built on AWS Bedrock. Building RAG pipelines, semantic search, knowledge ingestion, and vector DBs for LLM-powered compliance automation.
Education
Reva University
B.Tech
CGPA - 9.35
Skills
DSANLPDBMSRAG SystemsModel Fine-tuningDatabasesMLOpsVector DBsData StructuresPyTorchFAISSScikit-learnGitTensorFlowNLTKPandasNumpyChromaDBDockerGCPLangchainFirebaseREST API’sC++
PROJECTS
Neo: On-Device AI Assistant with Adaptive Learning
Built a fully offline Android AI assistant using quantized LLaMA via llama.cpp + JNI, achieving 4 ms inference latency with no network dependency. Integrated ONNX Runtime, Room DB, and voice (STT/TTS) for end-to-end interaction. Implemented persistent RAG pipeline using MiniLM embeddings and FAISS, enabling semantic memory recall across sessions entirely on-device.
ArtFlow: Neural Style Transfer
Implemented neural style transfer using VGG19 feature maps and Gram matrix style loss, achieving high-fidelity stylization in 300 iterations on GPU. Separated content and style representations from different network layers. Applied layer-wise weighting to control style intensity. Deployed as a web demo using HTML/CSS/JS for anyone to use.
Saathi: AI for Everyone
Built an AI voice assistant on Android targeting 300M+ low-literacy Indians, supporting 8 auto-detected Indian languages with sub-second STT/TTS response on mid-range devices. Shipped live screen assist using Android MediaProjection API floating voice orb with Llama 4 Scout vision, FLUX.1-schnell image generation, and real-time voice inference via Groq Llama 3.3 70B + Sarvam STT/TTS, all running zero cloud dependency for core features.
Engineered on-device RAG pipeline using ObjectBox vector DB, ONNX Runtime and MiniLM embeddings for persistent semantic memory.
CERTIFICATIONS
Programming in Modern C++
NPTEL
ACHIEVEMENTS
First Runner-up, The Bengaluru AI Hackathon
Sapthagiri NPS University