AI for your work. Built on ALTAI.
Products & services
- Yada for model access and data policy.
- isanagent for agents built around your task.
- Custom systems, developed with your team.
LOCAL-FIRST AI · PRIVACY BY DESIGN
On screen: the Altai range lit in neon, seen from afar.
Which data, which model?
On screen: the skyline forms a gate. A request card enters it and is read; the person’s name and ID number are masked and the request passes. A second request, with customer data, is blocked.
Yada applies your organisation’s rules to model requests: keep a request local, mask detected sensitive fields, or block it. We configure model access around your data policy.
Data handling rules
Configured checks apply allow, redact, local-only or block actions to detected personal data and secrets.
Inspectable policy decisions
Operators can see the rule, model and scan action applied to a request. Raw prompts are not logged by default.
Apply your rules to model requests.
One endpoint, defined model choices
OpenAI-compatible applications connect to a shared model endpoint. Operator rules decide which local or cloud model receives each request.
Your agent’s tools. Your operating rules.
On screen: the agent is a ring of digital clouds above the summit, with the model at its centre and task instructions, skills, tools, plugins, access and approval around it. One module turns into your tools, and a new cloud is added for your plugin.
isanagent is our customizable agent harness. We configure task instructions, skills, plugins and access to your tools, including the points where a person needs to approve an action.
Model and agent configuration
Configure model providers, agent instructions and specialist subagents. Subagents can have their own models and allowed tool lists.
Execution boundaries and controls
Configure tool permissions, working directories and command approval rules. Hooks can add checks before tool calls and logging after them.
Skills, plugins and tool connections
Define procedures as skills. Plugins bring together instructions, MCP tool connections, subagents and hooks.
An agent harness shaped around your work.
Multi-step task execution
The agent uses tool results to inform its next steps. Depending on configuration, it can delegate to subagents and run code locally, in Jupyter or over SSH.
Built alongside you. Measured in use.
FIELDWORK / DEVELOPMENT / MEASUREMENT
On screen: a section through the summit. Today’s method is drawn as a dashed line, measurement points appear along it, and the first prototype’s line rises above it (illustrative).
An ALTAI engineer works with your team, observes the task and builds the first experiment. We compare usage, error rates and completion time with the baseline before expanding the solution.
Establish the baseline.
Who will use it, how long does it take today, and where do errors occur? We record the answers and agree acceptance criteria before the first experiment.
Set the data boundary first.
Which information stays inside, which fields can be masked, and which services may be called? These rules guide model, tool and storage decisions.
Expand from evidence.
The embedded engineer completes bounded automations. Multi-system work, custom models or sustained development get a separately scoped team and budget.
Working with ALTAI
We work with the people doing the task: document the problem, inspect the data and test the critical assumption with a small prototype. ALTAI leads the engineering; you bring process knowledge and priorities.
A 90-day start, monthly field engineering, training or a dedicated project.
Our chemistry product: ChemAgent.
On screen: the question “How much oxygen does 100 g of α-pinene use as it burns?”. The model writes the molecule as SMILES, the tokens fold into its structure, its carbons are counted and the answer locks: 328.8 g O₂.
An example of our domain-specific model and agent development. It connects chemistry knowledge to scientific sources and calculation tools, and can be adapted to research and analysis tasks.
- ChemAgent (Domain-specific AI product for chemistry) Molecules, sources and calculations in one research workflow.
Work with molecular data
Resolve common chemical identifiers, retrieve PubChem and ChEMBL records, and calculate selected molecular properties with tools such as RDKit.
Run calculations and data analysis
Use Python, Pandas and DuckDB-based tools for chemical calculations, queries over CSV and JSONL data, and statistical analysis.
Our foundations. Your application.
Our foundational technologies
Yada governs model access and data policy; isanagent runs agents that use tools. We combine these foundations with our open-source components to develop models, agents and applications for your organisation.
From the model to the interface and your existing systems.
From model training to agent development
TRAINING / DEVELOPMENT / DEPLOYMENT
On screen: seen from above, the summit’s contour lines read as a fingerprint. A scan crosses it and four features lock on one by one, numbered like the services, then join into a match: made for you.
A maintenance assistant, Turkish document search or a multi-step analysis tool. We connect our foundations to the task’s data, tools and operating environment.
Task-specific models: Data preparation · Training · Evaluation
We develop small language models for your terminology and required outputs. AfterImage can support training-example preparation; model selection and training account for your hardware. Evaluation uses examples held out from training.
Agents & automation: isanagent · Your tools · Approval rules
We build task definitions, skills and plugins on isanagent. Together we define the data an agent can read, the tools it can call and the actions that need approval. We evaluate completion of the full task, as well as the answer.
AI software development: Interfaces · Business logic · Integrations
Document assistants for service teams, checking and reporting tools for operations, analysis interfaces for researchers. We build the software people use to inspect sources, correct outputs and continue the work in their existing systems.
Custom development & integration
Model & data infrastructure: Yada · Local deployment · Model serving
We prioritise running models within your environment, designing compute, serving and data flows together. Yada applies data rules to model calls. We test latency, memory use and operating cost against the intended workload.
Explore model access & data policy
Start with a single task, or build AI into software your team already uses.
Open source. Part of the same foundation.
SOFTWARE / MODELS / DATA
On screen: the range engraved on paper. The four projects’ summits are marked one by one with numbered rings.
Akana processes Turkish text, AfterImage produces training examples, letsearch provides retrieval, and llm-food prepares model-readable content. We connect these components to client systems and develop the adaptations they need.
Akana: LANGUAGE
Akana is a Turkish natural language processing toolkit built in Rust with Python bindings. Tokenization, morphology, normalization, readability, and text chunking provide a text-processing layer for search and language applications.
AfterImage: DATA
AfterImage creates conversational, tool-calling, structured-output, and preference datasets from task instructions and optional source documents. It provides the data-generation layer for teaching a model a specific task, with generated examples reviewed and evaluated before training.
letsearch: RETRIEVAL
letsearch provides a semantic search layer by indexing text collections as vectors. It imports JSONL, Parquet, or Hugging Face datasets, embeds selected fields, and exposes indexed collections through a search service.
llm-food: DOCUMENTS
llm-food converts supported documents and web pages into Markdown as a foundation for search, review, and model-input preparation. It is available through a FastAPI service, Python client, and command-line interface.
Inspect the code. Use it in your system. Build the integration with us.
Which task should we tackle together?
On screen: dawn over the range. A neon line traces the skyline, comes down and draws the “Talk to us” button.
Tell us about the work your team repeats, waits on or checks by hand. Together we’ll identify which foundations fit and what needs to be developed for your environment.