Llama
Qwen
The shape of a walkable city
Review article · 2025Street connectivity and daily access
↗ View source ⌑
If you can dream it, we can build it; every possible digital product delivered by professional AI engineers.
Our founders’ research is published at international AI conferences.
Explore selected research papers and original lab reports below.
Check out our researchYour ideas deserve to come to life, beautifully built down to the last detail.
Pick where you're stuck.
1 / 6. From your idea to a finished product.
What differentiates our innovations and services?
1 of 6: We trim the AI fat
What our services include
The latest AI models
State-of-the-art AI models, connected to your app. We use the API or CLI that fits the way you work.
AI and models
Business essentials
Media and play
Voice and messaging
Data and visualization
Security, finance and Web3
Pick how much finishing you need.
Finish it yourself
Lens Kit
Prompts written for your project, plus our pre-launch checklist.
- Prompts tailored to your project
- Pre-launch hardening checklist
- Works with the AI tools you already use
- Yours to keep
We check your finish
Clear Sky Kit
A real person tests your AI-built app before your users do.
- Every feature tested by hand
- Checked against the failures in our lab report
- Plain-language weakness report
- Fixes for what we find
We finish your site
Constellation Kit
Your website, planned with you and built for you.
- Planned with you in our chat
- Designed and built for you
- Every page and form tested by hand
- Clutter cut before launch
We finish your app
Telescope Kit
A custom app built around how you work.
- Interviewed until the plan is clear
- Built on apps we've already shipped
- Every feature tested by a real person
- Checked for launch problems
- Fixed price, no hourly billing
- Delivery date in writing
- Full refund if we can't take your project on
Stargazing Creativities Lab research
- SCL-2026-01Cutting the Fat: A Self-Applied Survey of Excess in AI-Generated Work
- SCL-2026-02Before You Ship: Failure Modes and Pre-Launch Checks for AI-Built Applications
Founders' published research
arXiv:2610.02190A* NeurIPS 2026Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-TuningWhen an AI learns, it improves in small steps, and picking the size of each step is tricky: too small is slow, too big throws it off. ZFO checks a couple of spots ahead before each step, which often helps the AI learn better.arXiv:2610.02199TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-TuningTeaching an existing AI a new skill normally takes a huge amount of computer memory. TACO cuts that memory so much that very large AI models can be retrained on a single graphics card, while staying about as accurate and fast.arXiv:2606.13767Beyond LoRA: Is Sparsity-Induced Adaptation Better?LoRA is a popular shortcut for customizing an AI without retraining the whole thing. This paper tests simpler, cheaper versions of that shortcut and finds they work about as well while using less time and memory.arXiv:2606.00024ART: Attention Run-time Termination for Efficient Large Language Model DecodingWhen an AI writes a reply to a long document, it keeps looking back over everything it has read. ART notices when looking back further won't change the answer and stops early, so replies come out up to 20% faster with the same quality.
About
Our Founder, Cristian McGee, is an AI researcher whose work on optimizing large language models and transformers has been accepted and published at conferences such as NeurIPS and Stars Library. He founded Stargazing Creativities to build ambitious software with the same rigor he brings to research.