Ravikumar M. N.Senior AI Engineer at C5iOpen to new roles
I build agentic AI systems that hold up in production: fast, governed, and auditable.
Multi-agent assistants, hybrid retrieval, and document intelligence on Databricks and Azure, from fine-tuning to CI/CD.
Impact in numbers
- <800msaverage response latencyMulti-agent assistant, concurrent multi-tenant sessions
- <120msvector search latencyInside a 1.8s end-to-end agentic RAG copilot
- −30%analyst query turnaroundNatural-language agents over governed enterprise data
- <3sdocument ingestion per pageDocling parsing and LLM extraction, with 1.5s grounded QA
Selected work
Systems I’ve designed and shipped
Each case study covers the problem, the constraints, the trade-offs, and the numbers.
- Read case study
C5i · 2026 – present
A governed multi-agent assistant for enterprise workflows
A LangGraph assistant that retrieves, reasons, and acts across business workflows, with sub-800ms average responses and an audit trail on every turn.
- average response latency
- <800ms
- end-to-end task time
- −22%
- analyst query turnaround
- −30%
- LangGraph
- Databricks Model Serving
- Unity Catalog
- Databricks Genie
- Azure DevOps
- Read case study
Project
An enterprise GenAI copilot with agentic RAG and tool use
One copilot over documents, tables, and analytics tools, with hybrid retrieval and query planning, answering in 1.8s end to end.
- end-to-end response latency
- 1.8s
- vector search latency at scale
- <120ms
- LangChain
- LlamaIndex
- Azure
- Read case study
Zyoin · Acranton · projects
Agentic document intelligence, from PDFs to structured data and answers
Docling parsing and LLM extraction turn PDFs, forms, and reports into JSON, a knowledge graph, and grounded QA, at under 3s per page.
- ingestion per page
- <3s
- grounded QA response
- 1.5s
- Docling
- LLMs
- Gemini API
- Read case study
Acranton · projects · 2022 – 2025
Model adaptation with LLMs, embeddings, and classifiers
Fine-tuning where prompting and retrieval run out, using SFT with LoRA, contrastive embeddings, and transformer classifiers taken from 78% to 89% accuracy.
- sentiment accuracy, up from 78%
- 89%
- inference latency
- −30%
- PyTorch
- Hugging Face Transformers
- PEFT / LoRA
Also built
Agentic research automation
−50% manual research effort
CrewAI agents that generate personas, run domain research, and compile structured reports.
CrewAI · LangChain
Conversational PDFs and websites
95% query success rate
Turns static PDFs and sites into queryable knowledge systems.
LangChain · LlamaIndex
Multimodal chatbot
+35% user engagement
Routes across three models to handle text and image interactions in one assistant.
LlamaIndex
Handwritten prescription → JSON
<5s per prescription
Structured extraction from handwritten medical prescriptions with Gemini Flash.
LangChain · Gemini
Write-up on the blog
Semantic search over NeurIPS 2022 papers with Sentence-BERT embeddings and an Annoy index.
Sentence-BERT · Annoy
Retinopathy diagnosis
+12% classification performance
Classifier for Central Serous Retinopathy.
PyTorch
Face-mask detection
93% precision
Real-time VGG19 classifier built during COVID-19.
PyTorch · VGG19
Bird-call classification
88% precision
Audio classification with signal detection.
PyTorch
Experience
Where the work happened
Shipping models to production since 2021: classical NLP and computer vision first, then LLM and agentic systems on enterprise data platforms.
Senior AI Engineer · C5i
Jun 2026 – Present
Full-time · India
Enterprise multi-agent assistant on LangGraph and Databricks, with governance built in.
- Architecting a multi-agent assistant on LangGraph (specialized agents for retrieval, reasoning, and task execution) that holds sub-800ms average response latency across concurrent multi-tenant sessions and cut end-to-end task time 22%.
- Serve agents on Databricks Model Serving with Unity Catalog lineage and access control on every agent-facing source; Genie-based agents cut analyst query turnaround 30%.
- Built governance in from day one (model versioning, prompt and response auditing, access policies) plus Azure DevOps CI/CD that cut release cycle time 20% and deployment failures 15%.
- LangGraph
- Databricks
- Unity Catalog
- Azure DevOps
AI Engineer · Zyoin
Nov 2025 – May 2026
Contract · India
Agentic RAG and synthetic data for enterprise analytics and customer intelligence.
- Designed and deployed an agentic RAG chatbot for enterprise knowledge, fed by Docling-based ingestion that turns unstructured documents into structured data.
- Built an agentic synthetic-persona generator for financial survey datasets to improve downstream model robustness and segmentation analysis.
- Ran the model lifecycle on Databricks, Azure, and MLflow: experiment tracking, versioning, and cloud deployment.
- RAG
- Docling
- Databricks
- MLflow
Machine Learning Engineer · Acranton
Jan 2022 – Aug 2025
Full-time · Davangere, India
NLP, computer vision, and generative AI, from prototype to production.
- Fine-tuned transformer NLP models from 78% to 89% sentiment accuracy while cutting inference latency 30%; custom classifiers raised pipeline accuracy 18% and cut processing time 25%.
- Built computer-vision pipelines for image classification and object detection with CNN and GAN models, reaching 92% accuracy.
- Designed a Gemini API–based extractor that turns unstructured healthcare documents into structured data in under 3 seconds, raising throughput 30%.
- PyTorch
- Hugging Face
- CNNs
- Gemini API
Generative AI Engineer · ResearchBrains
Feb 2021 – Dec 2021
Internship · Remote, India
Generative AI services on AWS.
- Built AI services on FastAPI and AWS (EC2, S3, SageMaker) that cut generative-model inference response time from 5s to 2s.
- Integrated modular AI microservices with GPT-powered APIs into enterprise workflows, cutting manual content-creation time 40%.
- FastAPI
- AWS (EC2, S3, SageMaker)
- OpenAI API
Education
B.Tech, Electrical Engineering · Central University of Karnataka
2018 – 2022
IEEE student member; started the Enigma AI Club.
Toolkit
What I build with
The tools behind the systems above, grouped by where they sit in the stack.
- LLMs & agents
- LangGraph
- LangChain
- LlamaIndex
- CrewAI
- Agentic RAG
- Hybrid retrieval
- Tool calling & SQL agents
- Databricks Genie
- Guardrails & audit logging
- Docling
- Training & inference
- PyTorch
- Hugging Face Transformers
- SFT
- PEFT / LoRA
- Sentence Transformers
- Contrastive learning
- scikit-learn
- Optuna
- Inference optimization
- Data & MLOps
- Databricks Model Serving
- Unity Catalog
- MLflow
- Weights & Biases
- TruLens
- Azure DevOps CI/CD
- Docker
- SonarQube
- Pinecone
- Weaviate
- Chroma
- Elasticsearch
- Languages & infra
- Python
- FastAPI
- Flask
- Pandas
- Streamlit
- AWS (EC2, S3, SageMaker)
- Azure
About
How I build AI systems
Most of my work sits between a model and a business process: agents that retrieve, reason, and act on enterprise data. The model call is rarely the hard part. The hard parts are the latency budget, the retrieval quality underneath every answer, and knowing when an answer is wrong. I treat evaluation, grounding, and audit trails as part of the system, not something added after launch.
I design for the constraints production will impose anyway: concurrent tenants, governed data, compliance reviews, release pipelines that have to stay green. I reach for fine-tuning when retrieval and prompting run out, not before. I started in classical NLP and computer vision, and that still shapes how I debug: look at the data first, then the model.
Coursework & certifications
- Finetuning Large Language ModelsDeepLearning.AI
- Building and Evaluating Advanced RAG ApplicationsDeepLearning.AI
- Deep Learning SpecializationCoursera
- Machine Learning SpecializationCoursera
Contact
Building agents, retrieval, or LLM infrastructure? Let’s talk.
Email is the fastest way to reach me. My résumé, GitHub, and LinkedIn are here too.
Open to senior AI engineering roles · Based in India