maharajahu.lab
INDEPENDENT AI ENGINEERINGLOCAL FIRST / OPEN SOURCE

Engineering
beyond the
default.

Big models. Your hardware.
I push local AI further with faster kernels,
better tools, and a lot of curiosity.

Enter the labView GitHub
BUILT BY MAHARAJAHU · DRIVEN BY CURIOSITY.
COMPUTE COREPROCEDURAL 3D STUDY
LOCAL INTELLIGENCEYOUR MACHINE. YOUR CONTROL.
Initializing 3D
SCROLL TO EXPLORE↓
REFERENCE TEST BENCHRTX 5090 / 32 GB VRAM

Interactive artwork.
Not live GPU telemetry.

THE TOOLKITCUDAllama.cppQwenClaude CodeWebMCPCODE → MEASURE → REPEAT

SELECTED PUBLIC WORK

Ideas are cheap.
Ship the evidence.

Practical experiments turned into tools you can inspect, reproduce, and build on.

Performance engineeringMIT

Same model.
Less waiting.

CUDA attention optimizations for Qwen3.8-27B Unsloth Q6 on the RTX 5090. Faster decode, with the same weights and context settings.

C++ / CUDAGGUFllama.cpp
Explore the kernels
PUBLISHED V5 RESULT
+49.8%

measured decode throughput

189,935 populated input tokensSingle RTX 5090 · Qwen 27B Q6
See the conditions, not just the number
Claude CodeDesktop + CLI
LOCAL BRIDGEMessages · tools · vision
QQwen 27BYour GPU
Local AI integrationApache 2.0

Your AI.
Your rules.

A lightweight bridge that brings local Qwen into Claude Code Desktop and CLI, with streaming, native reasoning, vision, and tool calls.

JavaScript · Anthropic API
Browser-native agentsApache 2.0

Power needs
a pilot.

ToolBraid connects browser tools across sites through an explainable execution graph, with explicit human approval before external changes.

WebMCP · Browser extension
All public work on GitHub

THE RECEIPTS

Proof.
Not promises.

A fast screenshot is easy. A reproducible comparison is the part I care about.

  • Same Qwen3.8-27B Unsloth Q6 weights
  • Same Q4 KV cache, MTP3, and sampler
  • GPU-resident inference on RTX 5090 32 GB
  • Median of three 512-token runs per profile
Read the benchmark methodology
DECODE THROUGHPUTtok/s · higher is better
189,935 actual populated input tokens
Baseline53.76 tok/s
Optimized V580.53 tok/s
+49.8%measured decode improvement

Text-only profile. This measures decode, not total request latency. The initial 190K prompt still took about 148 seconds to process.

What these results do — and don’t — prove +

The workload is a long-prefix synthetic test, not a completed coding benchmark. Six paired response streams and five deterministic attention outputs matched exactly on the recorded tests. That is not a universal guarantee of unchanged quality. Peak VRAM and initial prompt processing did not improve. Vision-loaded timing uses text input; image support was verified separately.

Inspect the raw evidence ↗

THE HUMAN BEHIND THE MACHINE

Independent.
Curious by default.

It started with
“what if?”
It usually does.

That question is usually where my projects start.

I’m an independent developer exploring what consumer hardware can do with better software. I work across local inference, GPU optimization, and agent tooling — testing ideas, breaking things in controlled experiments, and documenting what actually holds up.

Not every experiment wins. The useful ones become public code, measured results, and instructions someone else can follow. The goal is simple: more capable tools, with more control in the hands of the person using them.

Follow the experiments on X Browse my GitHub

INDEPENDENT WORK NEEDS INDEPENDENT SUPPORT

Fuel the next
“what if?”

Every coffee goes toward more experiments, more testing, and more things worth sharing.

Buy me a coffee