AI/ML Engineer · Richmond, VA
▍
YVYagyesh Vyas
AI systems that run on your own hardware — built to hold.
I design and ship AI systems that run on your own hardware — fine-tuning, retrieval-augmented generation, and security tooling with real numbers behind them. The demo is easy; the fourth day is the job.
I am an AI/ML engineer who builds tools that start working while the cloud is still negotiating.
Who's writing this
I'm an AI/ML engineer — I build systems that run on your hardware, not the cloud's. I got my B.E. in Computer Engineering in 2022, worked as a Data & AI Developer through 2025, and I'm finishing my M.S. in Computer Science in August 2026.
What I actually ship: fine-tuned models that fit in 4 GB of RAM, RAG pipelines that keep your data out of third-party APIs, and a security tool that checks other AI tools for leaks. The vibe is local-first; the numbers are public.
“The best AI is the one that works when you're offline”
— on-device inference, as a way of life
Below: how I spend my time — and how that time has been compounding since 2022.
B.E. Computer Engineering
Gujarat Technological University — Computer Engineering — First class with distinction
Data & AI Developer
MKL Management — data pipelines, dashboards, first LLM integrations
Three tools, open source
PersonalForge · VibeGuard · Resume Analyzer — 96★ · 11 forks
M.S. Computer Science
University of the Potomac — AI/ML specialization — GPA 3.88
Three things I shipped
VibeGuard
open sourceA friend shipped an app a chatbot had written for him. His API key was sitting in the bundle. He had no idea, and neither did the model. 693 rules and taint tracking that follows a variable across scopes — on your machine.
PersonalForge
open sourceI wanted a model that had read my own notes. Every path I found started with a GPU invoice, so I spent months getting it down to a free Colab session. Drop in documents, get a quantized GGUF.
Resume Analyzer
live appI was applying for jobs and kept hesitating before uploading my resume to sites that never say what they keep. So I wrote one where the honest answer is nothing.
What I reach for
Fine-tuning
Retrieval
Security
Shipping
What I reach for — because it works, not because it's trendy.
Six design rules and a few thousand hours of practice. The stack stays boring on purpose: the harder the problem, the fewer surprises I want between my head and the machine.
Teaching models
LLMs · Fine-tuningFinding the right passage
RetrievalReading code for trouble
SecurityMaking it usable
UI / UXWrangling data
DataGetting it out the door
ShipKnife-fight a tokenizer
Loss curves, warm-up rates, and everything else that happens before “it works” — from setup to the working checkpoint.
Tune on a budget
Zero-shot → instruct → fine-tune on 2M samples with LoRA and DPO, from Colab if it has to be.
Keep the paper honest
Read the repo, run the code, and find the claim the README leaves out.
Make it hold
Evals, taint tracing, and 693 rules for catching the leaks nobody announces.
Fit it in 4 GB
Quantization, GGUF, embedding models, streaming — the same model, on your hardware.
Leave receipts
96★, 11 forks, 12 clients, and a 3.88 GPA — the numbers are public for a reason.
The ML pipeline end to end
How I actually build RAG systems: from a messy pile of documents to answers grounded in your own data — every stage running on your hardware, nothing leaving your network.live pipeline
PDF · notes · code — cleaned, deduped, split
Local embeddings that run on CPU
Vector store + top-k rerank
RAG + LLM — grounded in retrieved context
A grounded answer — your data, your machine
Cleaning stages
dedup · normalize · split
Embeddings
BGE-M3, runs locally
Bytes sent to cloud
every stage on-device
Reranked context
feedback loop above
Credentials & Certifications
Six certifications across model-building, cloud, and container orchestration — IBM, AWS and the Cloud Native Computing Foundation.
IBM AI Engineering Professional Certificate
Machine Learning · PyTorch · Keras2024IBM Machine Learning Professional Certificate
Regression · Classification · Pipelines2024
AWS Certified Cloud Practitioner
Core services · Pricing · Security2024AWS Certified Machine Learning — Specialty
Data engineering · Modeling · Deployment2025
CNCF Kubernetes Fundamentals
Containers · Orchestration · Services2025
Certified Kubernetes Administrator (CKA)
Cluster ops · Networking · Storage2026Proof such as it is
GitHub stars
11 forks
AI clients
75 MCP tools wired
Certifications
IBM · AWS · CNCF
M.S. GPA
University of the Potomac
Interactive Shell
Type a command to explore the portfolio — no login, no API, everything local.
Tell me what you're building
Open to full-time roles from August 2026. I read every email and usually answer the same day. If you want to poke at the code first, that's fine too.
Resume
AI/ML Engineer — on-device inference, fine-tuning, security tooling, and systems that run without a cloud bill. Updated for 2026.
AI engineering, fine-tuning notes, and the unglamorous half of ML — evals, pipelines, and shipping tools people actually run.
GitHub
The code behind the craft — VibeGuard's rule engine, PersonalForge's fine-tune pipeline, and local-first tooling with real issues to prove it.
Thank You For Visiting
Every tool on this page runs on your own machine — no telemetry, no login, no keys. If you're building something that has to hold on the fourth day, my inbox is open.