I make systems more efficient. First on factory floors, now in models.
The first time my code moved something physical, it was a conveyor belt. YOLOv5 spotting metal and glass as they went past, Arduino arms pushing each piece into the right bin. It was a benchtop prototype and nothing about it was elegant. But it worked, and the interest stuck.
Now I train language models and build retrieval systems. What I care about most is the part that usually gets skipped: proving the thing actually works.
Looking for something specific? grounded · pretraining · SQL agent · the paper · background · stack
Most RAG demos answer everything you ask them. That sounds like a feature. It isn't. A system that always answers is a system that will confidently invent a number when the documents don't contain one, and you won't find out until someone acts on it.
grounded either cites its sources or declines. That's the design constraint, and everything else follows from it.
Retrieval is hybrid: BM25 for queries where the exact word matters, vector search for the ones where meaning matters, then a BGE reranker to sort what comes back. Evaluation runs on RAGAS and gates changes, so a retriever I think is better has to prove it before it ships. It runs fully offline through Ollama when the data can't leave the machine.
Built on LlamaIndex, Chroma and LiteLLM. The first domain profile is energy.
I trained GPT-2 124M from scratch on 10 billion tokens, on a single H100.
The first run was a reproduction. No changes, no cleverness, just getting a known result to land where it was supposed to. That step is easy to skip and expensive to skip, because if you modify an architecture without a baseline you trust, you can't attribute the difference to anything. You just have two numbers and a story.
Then I ran a modified variant of my own against it.
A natural-language-to-SQL agent over an 8-table schema, roughly 2M rows.
It matched hand-written reference queries on 24 of 30 held-out questions, graded on result-set equivalence rather than string match, at 2.1 seconds median latency. That grading choice matters: two queries can look nothing alike and still return the same correct answer, and scoring on text would have counted those as failures.
Six out of thirty is not a rounding error. That number is on this page because it's the honest one.
Sole-authored. It shows that a standard gradient-conflict diagnostic misreads auxiliary-loss interference at the output projection.
The short version: when a model trains on more than one objective at once, those objectives can pull in different directions. There's a common way to detect that, using the angle between gradients. The paper argues that at one specific place in the network, the output projection, that measurement tells you the wrong thing.
Experiments ran on GPT-2 124M and 350M, both trained from scratch on 500M tokens. Preprint forthcoming.
Center for Energy Research and Policy, Arizona State University. I built an LLM-assisted retrieval pipeline over more than a thousand energy-policy documents, which turned a literature review that took about a week into one that took days. I also shipped internal retrieval tools for researchers who don't write code, running open models locally so sensitive data never left the lab.
Green Ponds, India. An automation and robotics group. I joined as a data analyst and spent a year reconciling plant records by hand, feeling every pain point of doing it that way. Then I was promoted and got to build the systems that replaced them: the bridge pulling PLC and SCADA data into reporting databases over Modbus and OPC-UA, and monitoring for the robotic welding cells.
Getting to fix the exact problem you spent a year suffering is rarer than it should be.
Human Systems Engineering Lab, ASU. Data collection for a four-year study on how trust and distrust spread through human-AI teams, drawing on a UAV simulator and measures of team performance and communication.
Modeling · Python · PyTorch · scikit-learn · Hugging Face · computer vision (YOLOv5, OpenCV)
Retrieval and agents · RAG · LlamaIndex · Chroma · BM25 hybrid retrieval · RAGAS · LiteLLM · Ollama · Groq
Data and platform · SQL (MySQL, PostgreSQL) · NoSQL · Docker · FastAPI / Flask · git · C++
AWS Certified Generative AI Developer – Professional · AWS Certified Machine Learning Engineer – Associate