Enterprise RAG and semantic search
A grounded question-answering system spanning query understanding, expansion, dense and hybrid retrieval, reranking, evidence selection, citations and reflective retries.
AI / Machine Learning Software Engineer
I build reliable AI systems from retrieval to GPU kernels.
I’m Ranjith Vutnoor, an IIT Bhilai alumnus working across production RAG, LLM evaluation, agentic workflows, model fine-tuning and GPU-accelerated machine learning.
Open to thoughtful AI/ML opportunities
Hyderabad, India
2.5+ years in production AI systems
240+engineering work items delivered across features, reliability and defects
12.68×encoder speed-up in CUDA-accelerated thesis experiments
0.53→0.668Nemotron reasoning challenge score improvement
240+algorithm problems solved while strengthening interview fundamentals
Selected work / 01
Systems designed around evidence, evaluation and engineering constraints.
The public case studies below are intentionally sanitised. They show the technical approach without exposing confidential data, internal code or client-specific architecture.
EvalStudio — LLM quality workspace
An evaluation workflow for repeatability, golden-set regression, LLM-as-judge scoring and release-oriented quality gates.
GPU-accelerated anomaly detection
Extended MSCRED and ConvLSTM for industrial time-series anomalies, then accelerated convolution through im2col, GEMM, shared memory and a custom PyTorch CUDA extension.
Nemotron structured reasoning
Built compact synthetic reasoning traces, used a 4-bit QLoRA-style workflow, resolved model compatibility issues and improved benchmark accuracy.
Experience / 02
Building depth across the AI application stack.
I work best where research ideas must survive real data, production constraints and measurable quality expectations.
2023 — present
AI / Machine Learning Software Engineer
HCLTech · Enterprise AI systems
Developing and maintaining a production knowledge assistant for complex technical documentation.
- Engineered query classification, expansion, vector and hybrid retrieval, reranking and grounded response generation.
- Improved reliability through citation validation, guardrails, retry logic, observability and regression testing.
- Supported model migrations across multiple OpenAI model generations and structured response workflows.
- Designed incremental metadata-driven ingestion to avoid unnecessary full-index rebuilds.
2021 — 2023
M.Tech — Data Science and Artificial Intelligence
Indian Institute of Technology Bhilai
Built a research foundation across machine learning, deep learning, data science, optimisation and GPU computing.
- Completed a thesis on industrial multivariate time-series anomaly detection.
- Worked with MSCRED, ConvLSTM, PyTorch, CUDA, im2col and GEMM optimisation.
- Benchmarked CPU, global-memory GPU and shared-memory GPU implementations.
Independent work
Applied research and product experiments
Kaggle · Local LLMs · Desktop software
Exploring model adaptation, local inference, reasoning datasets and human-centred AI product ideas.
- Fine-tuned a LoRA adapter for the NVIDIA Nemotron reasoning challenge.
- Compared quantised language and vision-language models through local inference tools.
- Prototyped an Electron desktop companion with productivity and Pomodoro workflows.
Expertise / 03
A practical stack for modern AI engineering.
I care about the complete path from source data to user-facing reliability—not only the model call.
RAG and agents
Query understanding, expansion, embeddings, hybrid search, reranking, reflective retrieval, citations, structured outputs and tool integration.
Quality engineering
MRR, NDCG, hit rate, relevance labels, golden datasets, LLM-as-judge, repeatability, regression gates and failure analysis.
Fine-tuning and inference
LoRA, QLoRA-style workflows, PEFT, synthetic reasoning data, quantisation, checkpoint selection and local model evaluation.
GPU-accelerated ML
CUDA kernel programming, im2col, GEMM, shared-memory tiling, custom PyTorch extensions, ConvLSTM and performance benchmarking.
Application development
Python, SQL, PostgreSQL, APIs, React, Electron, Git, testing, data processing and production troubleshooting.
Cloud and observability
Azure OpenAI, Azure App Service, Key Vault, Application Insights, retries, diagnostics and operational feedback loops.
Research tooling
PyTorch, Hugging Face, PEFT, TRL, bitsandbytes, Qdrant, Jupyter, Ollama and LM Studio.
Algorithms and systems
Data structures, algorithmic problem-solving, system design, distributed-system concepts and model-serving architecture.
Writing / 04
Technical notes built from real implementation work.
Long-form articles are being prepared for this site and DEV Community. The outlines are available now.
In preparation
GPU-accelerating MSCRED with CUDA, im2col and GEMM
From convolution bottleneck to a custom PyTorch extension and shared-memory performance analysis.
In preparationFine-tuning Nemotron-3-Nano-30B for structured reasoning
Compact synthetic traces, 4-bit adaptation, compatibility fixes and deterministic evaluation.
In preparationEngineering an evaluation-driven enterprise RAG pipeline
Query expansion, hybrid retrieval, reranking, reflective search and measurable relevance.
Beyond work / 05
Curiosity beyond a single technology stack.
Product and technology exploration
I enjoy evaluating local AI models, personal computing, developer tools and product concepts—especially where engineering decisions must balance quality, latency, privacy and usability.
Travel and cultural exploration
Travel, temples and heritage destinations offer a useful counterweight to technical work. They encourage patience, observation and a longer view of systems and human behaviour.
Contact / 06
Let’s build something rigorous and useful.
I’m interested in AI/ML engineering roles and technical conversations involving production RAG, evaluation, LLM systems, model adaptation and GPU-accelerated machine learning.