Gen AI Solution Architect · Bengaluru
Building AI systems that hold up in production.
I'm Nagendra. I design and build copilots, RAG and agentic systems, ML platforms and the MLOps around them. These are the write-ups of what I built, what broke, and what I learned along the way.
What I write about
Featured projects
- Vision-language modelsComputer visionGenAI
Vision-language models in practice: from building one to retrieval-augmented visual QA
A hands-on deep dive into vision-language models: building a LLaVA-style VLM from scratch, comparing local and cloud VLMs on document extraction and visual question answering, and finding that in visual RAG the retriever, not the generator, is usually the bottleneck.
- AzureAWSLLMOps
Re-platforming intelligence: moving ML and LLM workloads from AWS to Azure
How we migrated a regulatory research platform's traditional ML models and LLM pipelines from AWS SageMaker onto a central Azure AI platform, using shadow deployments, an LLM gateway and separate migration tracks for ML and LLM workloads.
- CopilotsRAGAgents
Copilots as living systems: evolving an enterprise copilot into a multi-brain architecture
How an enterprise copilot grew from a single documentation RAG into an orchestrated set of specialised brains, and what we learned about retrieval, NL-to-SQL, evaluation and orchestration along the way.
More write-ups
Recommender systemsEvaluationRecommender systems in practice: ask how it was measured
A five-part deep dive into recommender systems on MovieLens and H&M: classical baselines, two-stage retrieval and ranking, cold start, whether deep learning pays off, and how the evaluation protocol alone can move accuracy 7–12× and reorder models.
- MLOpsForecastingTime series
From a handful of markets to global coverage: lessons from scaling demand forecasting
Scaling two hotel demand forecasting systems from a few launch countries to more than 50: why per-hotel forecasting broke and needed new lead-time features, why market-level forecasting scaled in two weeks, and the MLOps patterns that made adding a country a config change.
- AgentsAgentic AIGenAI
One platform, many agents: designing a hybrid enterprise AI architecture
A high-level enterprise AI architecture for an engineering and construction company: one shared platform that combines agents, retrieval and traditional ML, proven on three very different use cases.
RAGDocument AIGenAIBeyond text RAG: document intelligence for tenders with hundreds of pages, tables and CAD drawings
Three RAG experiments on a real 450-page EPC tender showed that document coverage and query intelligence, not the LLM, limit tender analysis. Multimodal ingestion, query decomposition and role-based retrieval close the gap.