→ QLoRA fine-tuner→ adyan.me→ vLLM inference server→ next project→ QLoRA fine-tuner→ adyan.me→ vLLM inference server→ next project→ QLoRA fine-tuner→ adyan.me→ vLLM inference server→ next project

scroll ↓

01 / WORK
Sunkiss Academy

Sunkiss Academy

ReactTypeScriptResponsiveSEOClient Work

Full production website for a licensed home-based preschool in Lynnwood, WA. Built and deployed for a real client — live at sunkissacademy.com. Features bilingual program pages, enrollment flow, parent reviews, meal schedule, and gallery. SEO-optimised with structured data, sitemap, and robots.txt.

view project →
Matchbox

Matchbox

ReactAWS AmplifyGraphQLDynamoDBTailwind

Web app that connects developers to projects matching their skills, interests, and values. Built a filtering algorithm that surfaces relevant projects from other developers. Six-page React frontend backed by AWS Amplify, GraphQL, and DynamoDB — deployed to production at appmatchbox.com.

view project →
In Progress

In Progress

TBD

Next project loading...

Python·TypeScript·Next.js·React·PyTorch·TRL / PEFT·vLLM·CUDA·Tailwind CSS·Framer Motion·Node.js·Git·Linux / WSL2·REST APIs·Python·TypeScript·Next.js·React·PyTorch·TRL / PEFT·vLLM·CUDA·Tailwind CSS·Framer Motion·Node.js·Git·Linux / WSL2·REST APIs·
Python·TypeScript·Next.js·React·PyTorch·TRL / PEFT·vLLM·CUDA·Tailwind CSS·Framer Motion·Node.js·Git·Linux / WSL2·REST APIs·Python·TypeScript·Next.js·React·PyTorch·TRL / PEFT·vLLM·CUDA·Tailwind CSS·Framer Motion·Node.js·Git·Linux / WSL2·REST APIs·
02 / ABOUT
“I build things that learn.”

0+

Years exp

ML+Web

Focus

Growing

Projects

I'm a software engineer and ML practitioner who builds on both sides of the stack. On the web side, I make fast, animated, opinionated interfaces. On the ML side, I fine-tune and deploy language models — close to the metal.

Right now I'm building out my project portfolio, training QLoRA adapters on consumer GPUs, and shipping things that work in production.

03 / ML & EXPERIMENTS
train.py

# QLoRA config — RTX 3080 10GB

model = "Qwen/Qwen2.5-Coder-7B-Instruct"

dataset = "iamtarun/python_code_instructions_18k"

lora_r = 16 # rank

bits = 4 # bnb 4-bit quant

→ 20 steps, loss: 1.42 → 0.87

QLoRA Fine-Tuner — Qwen2.5-Coder-7B

serve.sh

# vLLM 0.11.0 + Qwen2.5-Coder-14B-AWQ

python -m vllm.entrypoints.openai.api_server \

--model Qwen/Qwen2.5-Coder-14B-AWQ \

--quantization awq_marlin \

--gpu-memory-utilization 0.85

→ Served on RTX 3080, 8.86GB VRAM

vLLM Inference Server

04 / CONTACT

Let's work.

adyanzahid786@gmail.com →GitHub →LinkedIn →