# about/README.md · last edited just now

Fruitle

const handle = "fruitle"; // real name: Luca

Junior developer with the calloused hands of a senior. I build local-first AI, pry open old GPUs to make them useful again, and design storefronts that actually convert.

$ view source on GitHub → send email ⌘E case studies resume.pdf

role · independent engineer based · Cape Town, ZA focus · local AI · GPU · commerce status · taking work

YRS in commerce

0+

MCP servers shipped

0

GPUs revived

0

QUANTIZED models

0+

// skills.ts · what I do, how deep, what it costs you

The toolkit, honestly itemized.

02 · CAPABILITIES

01 · core

local_ai()

run frontier-class models on hardware you already own

GGUF / GPTQ / AWQ quantization
llama.cpp / vLLM / TensorRT-LLM backends
KV-cache + speculative decoding tuning
offline RAG with hybrid retrieval

02 · infra

mcp_servers()

protocol-grade tool servers for autonomous agents

Model Context Protocol (stdio + SSE)
tool surfaces over filesystem, db, browser
sandboxing & capability scoping
agent loops with budget & reflection

03 · hardware

gpu_kernels()

squeezing AI workloads out of cards everyone gave up on

VBIOS reflashing, clock + voltage tables
CUDA / ROCm / Vulkan compute kernels
memory-bandwidth aware op fusion
P40 / M40 / MI50 cluster bring-up

04 · commerce

storefronts()

9 years building layouts people actually buy from

Shopify · Woo · headless commerce
CRO-driven layout systems
Core Web Vitals & PSI ≥ 95
checkout funnels & A/B harnesses

05 · agents

autonomous_loops()

agents that finish jobs without supervision

plan / act / critique architectures
tool routing + cost governors
long-running task scheduling
vector memory + episodic recall

06 · perf

backend_tuning()

when "it's slow" needs a real number attached

flame graphs, eBPF, perf, nsight
Node / Python / Rust async perf
Postgres / Redis hot-path tuning
queue + worker topologies

stack.json json

{
  "languages": ["TypeScript", "Python", "Rust", "C++", "CUDA", "Go"],
  "runtimes":  ["llama.cpp", "vLLM", "TensorRT-LLM", "ONNX"],
  "frameworks":["Next.js", "Astro", "Hydrogen", "FastAPI", "Actix"],
  "hardware":  ["Tesla P40", "M40 24G", "MI50", "RX 580", "3090"],
  "motto":     "if a senior could ship it on a budget, so can I."
}

// projects/ · selected work across three problem domains

Three threads I keep pulling on, braided together.

03 · WORK

01 · local-ai.md

Frontier-class models, on your hardware.

Quantization pipelines, custom MCP servers, and agent loops that don't require sending a single token to a third party. Built for teams that have data they can't ship and compute they already paid for.

GGUF Q4_K_M / Q5_K_S
vLLM + speculative
MCP stdio & SSE
autonomous agents

read the case study →

02 · gpu-vbios.yml

Old silicon, refused to retire.

My ongoing project: turning datacenter-pull GPUs into legitimate AI accelerators. Custom VBIOS work, kernel-level optimization, and undocumented quirks documented — so a kid in Soweto can run a 13B model on a $80 card.

VBIOS reflash
CUDA kernel fusion
thermal headroom
P40 / M40 / MI50

read the technical log →

core clk

1531 MHz

mem clk

3615 MHz

vram

10.1 GB

tdp

137 W

temp

62 °C

03 · ecommerce.tsx

Storefronts that actually close the sale.

Nine years of Shopify, headless commerce, and CRO. Layouts engineered around the way a hand actually moves on a phone, not the way a Figma file looks on a 27" monitor. PSI ≥ 95, AOV up to the right.

Shopify · Hydrogen
Headless commerce
CRO experiments
PSI / CWV tuning

see the work →

// experience.log · tail -f, ascending

The trajectory, unfiltered.

04 · LOG

2026.05 SHIP Released VBIOS toolkit v0.3 — flashing & clock-table editor for Pascal-era datacenter cards. +27 cards revived in field.

2026.02 OK Built autonomous research agent for a fintech client — runs nightly, drafts memos, files them. Cost: $0.04/run on local llama-3.

2025.11 INFO Started GPU revival project — making "AI accessible" not as marketing, but as a 24-GB card under R1500.

2025.08 SHIP MCP server suite — filesystem, browser, postgres, calendar. Wired into Claude Desktop and a custom agent runner.

2025.03 OK Quantization pipeline — automated GGUF/AWQ conversion + eval harness. 60+ models pushed through.

2024.09 WARN Pivot — full-time commerce work narrows; AI infra requests overtake. Decision: lean in.

2023 INFO Headless commerce migrations — Shopify → Hydrogen for two SME clients; PSI 60 → 96.

2021 SHIP First million-rand store built end-to-end — checkout, fulfilment, content, theme.

2019 OK Freelance, full-time — Shopify themes, WooCommerce, Liquid & PHP archeology.

2017 INFO First paid e-commerce build — custom theme, 13 years old, $50, learned more than any textbook.

# contact.sh · pick a channel; I answer fast

Let's build something difficult.

05 · CONTACT

Got a project that needs real engineering?

Local-first AI deployments, GPU bring-up, an e-commerce stack that won't fall over on Black Friday — happy to scope a call. Junior rates, senior delivery.

$ email Luca@smallroomproductions.online → $ phone +27 84 335 7998 → $ web smallroomproductions.online ↗ $ github /fruitle ↗ $ call book a 20-min intro →

ACCEPTING WORK · Q3 2026

"I'd rather ship a slow car that runs forever than a fast one that won't start tomorrow."

SIGNED · FRUITLE CPT · ZA