Headline

AI/ML Engineer | LLM Deployment · RAG Systems · Agents | Python · vLLM · Qdrant | Open to full-time roles

About

AI/ML Engineer with 1+ year building LLM systems at Synchronoss Technologies. What I've worked on: → Migrated LLM inference from Ollama to vLLM — benchmarked 16 concurrent requests vs Ollama's bottleneck → Engineered a RAG pipeline over 7M+ code chunks using Qdrant, BGE-large embeddings and AST-aware chunking → Built PyTestGen — AI unit test generator using local LLaMA (https://github.com/PrateekJanagouda/pytestgen) → Built MiniRAG — complete RAG API from scratch, no LangChain (https://github.com/PrateekJanagouda/minirag) Passionate about agents, RAG systems and LLM inference optimization. Open to full-time ML/AI Engineer roles in Bangalore. GitHub: https://github.com/PrateekJanagouda

Location

Bangalore Urban

AI stack

claude code

claude

codex

deepmind

gemini

hugging face

Work experience

Software Engineer Trainee (machine learning) at Synchronoss Technologies, Nov 2025 - Present · 1 yr, Bengaluru, Karnataka, India, Working on production LLM systems — retrieval, agents, and inference. • Built an internal agentic RAG assistant over a large Java microservice codebase. It answers plain-English engineering questions and returns answers backed by citations to the source, and is used by the engineering team day to day. • Designed it around two stores rather than one: a relational store for exact structural facts (symbols, call graph, service boundaries) and a vector store for semantic search. Structural questions like "what calls this" are graph queries — similarity search answers those confidently and wrongly. • Combined code-specialised and text-specialised embedding models with a cross-encoder reranker, and added a bytecode-analysis pass on top of tree-sitter so the call graph resolves across repository boundaries, which source parsing alone cannot do. • Built the evaluation harness: it generates questions from the index itself and scores answers against cited evidence, with a citation validator and confidence scoring, so retrieval quality is measured rather than assumed. • Built a second agent for configuration management that constrains the model to emit only RFC 6902 JSON Patch operations against a fixed schema, grounded in a real diff, instead of letting it rewrite YAML freely. Every generated change is schema-validated and dry-run before it can be applied, with rollback. Tech: Python, FastAPI, vLLM, Qdrant, PostgreSQL, tree-sitter, Helm, Docker

Machine Learning Intern at Synchronoss Technologies, Oct 2024 - Oct 2025 · 1 yr 1 mo, Bengaluru, • Evaluated open-source LLMs for Java test generation, then load-tested the serving layer from 1 to 2,000 concurrent users. Ollama dropped every request at 20 concurrent; moving serving to vLLM held 1,000 concurrent at 100% success and ~790 tok/s using continuous batching. • Benchmarked open-source code models (Llama, StarCoder, CodeGen) for JUnit test generation — set each up locally and evaluated output quality to pick the best performer. Integrated the Continue plugin with IntelliJ so developers could generate tests inside their editor. • Researched and benchmarked open-source face-recognition models, selected FaceNet, and integrated it through the DeepFace library, tuning it for face-clustering performance. • Built a media-processing pipeline that clusters faces across video and classifies expressions in real time using DeepFace's emotion-detection module. Tech: Python, Llama, StarCoder, CodeGen, vLLM, Ollama, DeepFace, FaceNet, Continue, IntelliJ

Education

Bachelor of Engineering - BE, Artificial Intelligence and data science, Visvesvaraya Technological University

mahesh pu college

Public links

github: https://github.com/PrateekJanagouda

prateekjanagouda.github.io: https://prateekjanagouda.github.io/