Independent AI Efficiency Lab
AI that costs less to run.So everyone can use it.
We design deep learning architectures that are efficient by design, then engineer how they run, from GPU kernels to serving, to cut both cost and latency. Lower cost is how AI reaches more people.
Selected work
Products, tools, and investigations at different stages of development.
Developer Platform
A shared console for accessing Blacktile's API services, managing keys, and understanding usage.
Early accessDocument Text Extraction
Text and layout extraction for real-world documents, with Thai and mixed-script content in focus.
AlphaInference Research
Investigating GPU kernels, memory movement, and state reuse in model inference.
ResearchLab Tools
Small tools for profiling, dataset generation, and repeatable experiments.
Research
Research areas
Questions we explore across computing systems, model inference, documents, and infrastructure.
Computational efficiency
When does moving data cost more than computing on it?
Understanding where computation, memory movement, and scheduling limit performance.
Inference systems
When is moving KV cache cheaper than recomputing it?
Studying the movement, storage, and reuse of model state during inference.
Document processing
How well does synthetic data transfer to real Thai documents?
Extracting text and layout from real documents, including Thai and mixed-script material.
Developer infrastructure
How can usage be metered correctly across retries and failures?
Designing small, reliable systems for access, metering, and product delivery.
Engineering philosophy
The principles behind our experiments and the software we build.
Measure first
We test assumptions and measure behavior before deciding what to optimize.
Keep complexity deliberate
Add dependencies and services when their value outweighs their cost.
Be precise about claims
Distinguish open questions, prototypes, measured results, and available products.
Design for failure
Retries, timeouts, and partial failures are part of the system, not afterthoughts.
Projects
Local LLM inference, optimized for cost.
We set up language models on your own hardware and tune them to serve more requests per GPU: quantization, batching, and KV-cache settings, measured before and after. Your data stays inside your network.