Yash Raj Pandey
I'm an AI systems engineer at UF IFAS, where I build agent and retrieval systems that run on university hardware. On my own machines I reproduce new model research, and I send fixes upstream when I find a bug.
I lead the AI agents work at UF IFAS. Generation runs on university GPUs, and retrieval stays on lab hardware because I benchmarked both and local won. I also took over Blue Omics, the genomics platform for the UF blueberry breeding program, when its lead developer left in August 2025. I publish failed reproductions too, and I correct a post when my own claim turns out wrong.
Open source
llama.cpp #24305, merged
Fix wrong RMS norm gradients on the CPU backend
On the CPU backend, the RMS norm backward pass returned a wrong gradient when the scheduler reused an input's buffer for the output.
LiteLLM #30652, merged
Fix Perplexity searches billed at 1/1000 of the price
When Perplexity did not return its own cost, LiteLLM billed search queries at 1/1000 of the real price.
Google TabFM #42, merged
Fix predict crashing on machines with two or more GPUs
With default settings, the first predict call crashed on any machine with two or more GPUs unless you pinned it to one device.
Since May 2026, 129 of my pull requests have been merged across 34 open source projects, and almost all of them are bug fixes.
Blue Omics
Research platform, 2025-2026
Genomics research platform for the UF blueberry breeding program
Joined at 81 source files , then 687, and the lead seat
Podium
Developer tool, 2026
Verified delegation for coding agents, where a shell command decides whether the work landed
The bot reports done , then the runner decides
Looma
AI infrastructure, 2026
Local-first project memory for coding agents
Recall@3 0.62 , then 1.00, on 0 third-party dependencies
mddocs
Collaboration tool, 2026
Git-native collaborative Markdown, with an agent API
An agent suggests , then a person accepts or rejects, in git
ApplyScore
Browser extension, 2026
AI resume gap-analysis extension
Postings hide in Shadow DOM , then 8 extractors, 1 fallback
Things I evaluate
I test new models and inference stacks on my own hardware and publish what I measure.
willitcall
Conformance suite + public matrix, 2026
The caniuse of local-model tool calling
50 tool-call scenarios , then 32 rows, 3 servers
TabFM Evaluation
Model evaluation, 2026
Independent evaluation that surfaced and fixed a multi-GPU crash in Google's tabular foundation model
Crashed on multi-GPU , then fix merged by Google Research
Writing
I would start with these two posts. In one I got a claim wrong, and in the other I lost two competitions the same way.
All writingTurboQuant on an M1 Pro - needle retrieval from 0% to 100% at 16K on MLX, and four bugs fixed in two llama.cpp forks
Called it a bug fix , then variance reduction, corrected
I named the question I failed to ask, then entered another competition and failed to ask it again. What actually wins these races, and why I keep missing it.
2 competitions entered , then 2 lost the same way
Football
I follow seven football competitions, so I built a dashboard for them behind a caching proxy. I also built a prediction pool and ran it through the 2026 World Cup.
World Cup 2026 Picks
Product, 2026
Self-hostable prediction pool for small groups
Skipped picks count as wrong , then no hiding in safe games
Football Hub
Web application, 2025