Teaching came naturally.
Long before AI systems and production pipelines, the signal was simple. Explaining hard things clicked. The classroom instinct has stayed constant.
AI systems - learning loops - human adoption
Sindhu Ghanta, Ph.D.
Teaching has been the constant. I build AI systems from scratch because the inside view makes them teachable. You see where the model breaks. You see where the abstraction leaks. You see where someone is likely to get stuck.
Operating principle
Build the system. Then teach from inside it.
The clearest lessons come from things people can inspect, question, and change. Prototypes, experiments, and workflows reveal where real understanding gets stuck.
The through-line
Teaching came first. Research gave it depth. Industry made it real.
Research made uncertainty useful.
A Ph.D. started as a path toward teaching. It became a love of experiments. I liked the humbling work of designing something without knowing if it will hold.
Industry made it concrete.
Building ML models, distributed systems, and early MLOps clarified the real question. How do you make what is being built understandable enough to use?
GitHub projects
A few projects I have been building and sharing.
Here are a few public projects from my GitHub. They show the kinds of AI systems, teaching tools, and technical explainers I like to build.
PyTorch - transformers - education
TinyGPT
A GPT model built from scratch in readable PyTorch. It covers BPE tokenization, causal self-attention, training, sampling, and perplexity evaluation.
Open repositoryMCP - dataframe QA - safety
MCP DataFrame QA
A local, read-only analytics server where an LLM proposes typed analysis plans, validates them, and executes safe dataframe operations instead of seeing raw data.
Open repositoryXGBoost - notebooks - visual explainer
XGBoost From Scratch
A numbers-first walkthrough of residuals, similarity scores, gain, learning rate, gamma, and lambda, worked by hand and checked against the real library.
Open repositorySchovia
Head of Machine Learning at Schovia.
I create explainer videos and practical lessons that make AI concepts feel intuitive. The work is to help people understand what is happening inside the system, not just memorize the tools.
Watch Schovia videosChunking, hybrid search, reranking, and query rewriting.
RAG quality as an engineering pipeline, not a single vector search checkbox.
Accuracy, faithfulness, hallucinations, and refusal behavior.
Answers have to be correct, grounded, and safe enough to use.
Prompt caching, long-context limits, and cleaner context design.
More tokens can help, but clean context and decision structure often matter more.
Proof, not polish
Receipts from systems that shipped, taught, and scaled.
countries reached
AI enablement programs for developers, product managers, and executives, with NPS 60+ and curriculum updates shipped within days of major model releases.
AI education videos
Technical explainers on RAG, LLM evaluation, vector search, MCP, transformers, and developer workflows.
granted US patents
Production ML interpretability and dataset suitability, plus additional pending work.
citations
h-index 11, with venues spanning Genome Biology, IEEE, ICMLA, USENIX HotEdge, and SPIE.
Artifacts I care about
Not just talks. Working systems with taste, constraints, and feedback.
From-scratch systems as teaching instruments
TinyGPT and XGBoost From Scratch turn algorithms into inspectable code, notebooks, and visual explainers instead of treating APIs as the lesson.
Safe local tools for real workflows
MCP DataFrame QA applies a research-informed pattern to local analytics. It uses compact schema context, typed plans, validation, capped outputs, and auditable execution.
Early MLOps before it had a clean label
Model monitoring, analytics platforms, feature workflows, interpretability, and dataset suitability systems at Parallel Machines, later acquired by DataRobot.
Clinical AI and computational biology roots
ML and computer vision pipelines for Harvard Medical School clinical research, including co-first-author work in Genome Biology.
Teaching experience
Insights from teaching AI every week.
The obvious is usually not obvious.
Expert intuition hides steps. Good teaching notices the missing step and makes it visible.
Uncertainty is part of the lesson.
Research taste means knowing what you do not know yet, then designing a way to find out.
Curriculum is infrastructure.
If the knowledge cannot update as the system changes, it is documentation theater.
Current coordinates