class YashRajPandey:
    role        = "AI Agents Architect @ UF IFAS"
    based_in    = "Gainesville, Florida"
    focus       = ["local-first LLMs", "RAG", "vector search", "AI agents"]
    also_builds = "full-stack production systems"
    philosophy  = "Understand the real problem. Ship the simplest thing that works. Measure. Iterate."

I build AI infrastructure and production software. I'm the AI Agents Architect at the University of Florida's Institute of Food and Agricultural Sciences, where I lead a function I proposed myself: self-hosted, air-gapped AI systems built on open-weight models.

The work is end to end: open-weight models served on-prem, retrieval that pairs dense vector search with reranking, agents that call real tools, and evaluation gates that decide what ships. None of it leaves the building, and it holds up under real users rather than a demo.

I joined UF as a Software Engineer in March 2025, was promoted to Lead Software Engineer in October 2025, and moved into the AI Agents Architect role in April 2026.

Independent evaluation of TurboQuant (arXiv 2504.19874), a near-optimal LLM quantization method, ported to run CPU-only on Apple Silicon. Fixed five blocking bugs, then benchmarked long-context retrieval across an MLX path and a llama.cpp Metal path.

Needle-in-a-haystack retrieval: 0% to 100% at 16K tokens, with a large KV cache memory reduction, all on a consumer M1 Pro with no dedicated GPU.

View the repo ->

A git-native, self-hostable editor for Markdown: real-time multiplayer, inline comments, and accept/reject suggestions, plus a first-class HTTP API so AI agents read, edit, and review documents the same way people do. Published on npm and built on proof-sdk.

• Local-first and git-native: every change is a commit, no central database to run
• Real-time collaboration with comments and accept/reject suggestion review, backed by a CRDT (Yjs) document model that merges concurrent edits without conflicts
• Agent-facing HTTP API with per-agent tokens and rate limits, so automated writers are first-class collaborators, not bolt-ons

View the repo ->

A Django, React, and PostgreSQL platform I designed and built from scratch for a research lab. It grew from zero to 5M+ live records across 32 data models and 58 API endpoints, replaced years of manual spreadsheet workflows, and became the primary system for every lab submission.

• Tuned PostgreSQL with explicit indexing and caching to hold low-millisecond median latency under concurrent access by 30+ researchers
• Built 7 ingestion pipelines for heterogeneous formats: PDF, Excel, CSV, Word, PowerPoint
• Automated R Markdown reporting, cutting a 2-3 hour manual process to 15-20 minutes
• Deployed on GCP with Kubernetes and Terraform; optimized the frontend from 8s to 3s load time

A published Chrome extension that scores how well a resume matches any job posting on the web, with evidence-linked gaps and no hallucinated fluff.

• Universal scraper that pierces Shadow DOM to read postings across LinkedIn, Greenhouse, Ashby, Lever, Workday and more
• Strict, evidence-based analysis: a confidence-weighted fit score with requirement-by-requirement matches linked to the exact resume bullets that prove them
• Privacy-first and bring-your-own-key (OpenAI, Anthropic, or Google), so data and model choice stay with the user

View on the Chrome Web Store ->

I fix real bugs in the AI infrastructure I build on: inference engines, ML frameworks, and the tooling around them. 30+ merged pull requests across 20+ open-source projects, each one a root-caused fix backed by a regression test.

LLM inference and ML frameworks

• llama.cpp (ggml) - fixed rms_norm_back producing wrong output under in-place aliasing on the CPU backend
• MLX (Apple) - signed-integer overflow in roll and tile; undefined behavior in arange(step=0)
• MLX-LM (Apple) - server 404 on short prompts; sampler top_k bound fix
• RAGFlow - broken language detection; Excel parser off-by-one chunking
• mem0 - Redis, FAISS, and Weaviate crashes and filtered-search truncation

Developer tooling

Cost-accuracy and edge-case fixes across a suite of open-source CLI and macOS tools: pricing-unit bugs (per-token vs per-request), timezone and Unicode handling, performance regressions, and more.

See all merged contributions ->

Herbert Wertheim College of Engineering Achievement Award, University of Florida. Merit scholarship awarded for top academic standing.

• M.S. Computer and Information Science and Engineering, University of Florida - GPA 3.8 / 4.0
• Semester Exchange, University of Florida - GPA 3.7 / 4.0
• B.Tech Computer Science and Engineering, Jaypee University of Engineering and Technology - GPA 9.1 / 10.0

I keep playbooks on what I learn shipping local AI, over on yashrajpandey.com:

• Self-hosting open-weight LLMs without sending data to a cloud API
• RAG that holds up in production: retrieval, reranking, and the evals that keep it honest
• Evaluation-gated releases for LLM systems

Off the clock: football (I built Football Hub and World Cup 2026 Picks because I love the game), tactical FPS, story-rich RPGs, and lo-fi for flow state.

Open to good conversations on AI infrastructure and local-first LLM systems.