🔭 A space to build, experiment, and learn together about AI systems.
🌱 I share hands-on projects, evaluation workflows, and lessons from exploring how AI systems work—and what it takes to make them useful. These projects are starting points for others to explore, reproduce, question, and extend.
👯 I work at the intersection of AI systems, developer platforms, and product decisions. My work spans open foundation models at NVIDIA and, previously, AI developer platforms and ecosystems at Intel. Explore the projects, share questions, and learn along with me.
What does it cost for a coding agent to produce an accepted patch?
A coding-agent workflow and tutorial measuring task success, context use, and inference cost on repository-maintenance tasks.
Does better retrieval lead to better answers—and what does reranking cost in latency?
Experiments evaluating retrieval quality, generated answers, and latency.
How can developers optimize local AI applications and what are the tradeoffs across inference stacks?
Hands-on lab demonstrating model serving, latency, and deployment tradeoffs across llama.cpp, vLLM, SGLang, and NVIDIA NIM, including techniques such as quantization and prefix caching.
- AI evaluation: building trust in agentic AI through evaluation of business outcomes.
- Contributor to docx4j, an open-source library for processing document content and structure in XML-based formats. I am acknowledged as Manimala Kumar in the docx4j 3.0.1 release metadata.
- Autonomous vehicle report processing: Documentation and examples from an OCR/NLP pipeline for California DMV collision reports. The current repository contains the write-up; pipeline source is not included.
- Clustering and feature transformation: Earlier machine-learning experiments with clustering, dimensionality reduction, and revenue-prediction data.
- Regression and feature engineering: Earlier learning exercises in price prediction.