I build LLM and RAG systems that turn messy, unstructured documents into clean data people can use. These days that means research papers, but the problem shows up everywhere: somewhere there's a person copying fields out of a PDF by hand, and usually a model can do most of it.
I'm currently an Applied AI Engineer at MindSpec, working on automated extraction from scientific literature. Before this I did an MS in Business Analytics & Data Science, and before that I spent a couple of years writing test automation at TCS. So I tend to think about two things at once: how to build the thing, and whether the output is reliable enough that someone will trust it.
At MindSpec I'm building a production LLM extraction pipeline that reads long neuroscience papers, dense tables and figures included, and turns them into schema-validated structured records. The interesting problems are the unglamorous ones: parsing tables and figures without garbling them, validating output against a schema that keeps changing, and keeping records deduplicated across 15+ documents. Curating a paper by hand used to take about three hours; the pipeline gets it to about 45 minutes plus a reviewer's sign-off.
Stack: LangChain, ChromaDB and PgVector, Docling, LlamaCloud, Pydantic, and Anthropic API.
- GenAI / ML: LangChain, LangSmith, RAG, vector databases (ChromaDB, PgVector), Docling, LlamaCloud, Pydantic, scikit-learn, TensorFlow, NLP
- Languages & data: Python, SQL, R, REST APIs
- Databases & warehousing: PostgreSQL/pgvector, MySQL
- Analytics & BI: Power BI, Tableau, Plotly Dash, Excel (Power Query, Solver)
- Other: Selenium, Git, Jupyter, JIRA
hr-agentic-assist — An agentic HR assistant that takes plain-language instructions and runs the onboarding busywork that normally lives in email threads and spreadsheets.
ecommerce-qa-chatbot — A shopping assistant that translates natural-language questions into SQL, answers product FAQs, and uses semantic routing to work out which kind of question you're asking.
Machine-Learning-Web-App-using-Streamlit — A binary-classification app built around the decision rather than the accuracy score. It surfaces the false-positive/false-negative trade-off so you can set thresholds sensibly when one kind of mistake costs more than the other.
There's more on my repositories page, including an LLM news-research tool with source attribution, plus data-engineering and BI work (ETL across 1,000+ tables, dashboards spanning finance, sales, and supply chain).
- 🎓 MS, Business Analytics & Data Science — Oklahoma State University
- ⚙️ B.Tech, Electronics & Communication Engineering — SRM University
- 🧪 Earlier: Quality Assurance at Tata Consultancy Services (Python + Selenium)
I'm open to roles in GenAI / ML engineering and data science. Easiest ways to find me: