We build AI that works alongside people.
AS Core AI develops agentic software for engineers and trains compact models to understand how useful, controllable AI can be built from first principles.
What we build
Intelligence becomes useful when it can act, explain its work and remain under human control.
Our product and research share one direction: turn model capability into practical systems without hiding what happens between a prompt and a result.
Our product
Ascora ADE
A cross-platform agentic development environment. Describe a task; the agent can inspect the project, edit files, run tools and show exactly what changed.
Reviewable changes
The agent reads, plans, edits and runs commands — then presents a diff before writing to disk.
Local when you want it
Use LM Studio or Ollama fully offline, or connect CLI, API and web-chat backends per task.
Reusable agent workflows
Blueprints connect multiple named agents, conditions, repeat loops, commands and delivery steps.
Experimental model training
We train small models to learn what matters.
Ascora Nano R1 is a from-scratch model project focused on Russian, English and developer tasks.
The work is experimental: it helps us study data pipelines, reasoning formats, evaluation and local inference. The long-term goal is to bring proven capabilities into Ascora ADE — not to present an unfinished research model as a production service.
Explore our public workAscora Nano R1
Prepare the data
Russian, English, code and structured technical documentation pass through cleaning, deduplication and packing.
Train from scratch
A compact decoder-only Transformer is pretrained, instruction-tuned and then reasoning-tuned in pure PyTorch.
Evaluate behaviour
Reproducible benchmarks measure language, code, arithmetic, reasoning quality and generation speed.
Earn product use
Only capabilities that become genuinely useful and reliable are candidates for future Ascora ADE integration.
How we think
Three principles connect the model lab to the product.
Keep control close
Local models and offline workflows should remain first-class choices wherever the task allows it.
Make actions legible
Reasoning, tool calls and file changes should be inspectable instead of disappearing behind a result.
Research must earn its way in
Experiments become product features only when they are useful, testable and reliable enough.
Start with the product