Welcome to Pytama
We design and build AI systems for businesses that run on operations, documents and data. Our work covers agents that carry out multi-step work, knowledge that improves with use, analytics you can question in plain language, and local models with observability built into the workflows you already have.
What we build
Systems that take on the repetitive and information-heavy parts of how a business runs, built to fit the tools and rules already in place.
Agentic workflows
Multi-step agents built as stateful workflows, with tool calling, persisted state and human checkpoints where a person needs to approve a step. Outputs are checked against your own taxonomies and governance rules.
Responsive operations
Systems that react to live signals from your operations, with every decision logged, reviewable and kept under your control.
Improvement pipelines
Feedback, evaluation and retraining loops that make your AI systems better over time, with clear visibility into what changed and who approved it.
Conversational analytics
Natural-language questions answered with SQL-grounded results, alongside dashboards for KPIs, trends and geospatial views of the data.
Local and open-weight models
Models run on your own infrastructure and wired into existing workflows, with confidence signals read from the model's internal state to flag answers that should not go straight through.
GUI and computer-use agents
Intent recognition across the stages of graphical task execution. Evaluated across benchmarks, where it outperformed chain-of-thought baselines.
Custom confidence metrics for your data
We build confidence metrics tailored to your data and your local models, so you can see which answers to trust and which to check. The example below is one scenario we studied, using Qwen3-8B and Llama-3.2-3B on survey data. It shows how a metric read from the model's internal state can outperform its own output confidence.
| Survey | Model | Output confidence | Hidden-state probe |
|---|---|---|---|
| OpinionQA (US) | Llama-3.2-3B | 0.674 | 0.676 |
| OpinionQA (US) | Qwen3-8B | 0.753 | 0.708 |
| Pew India | Llama-3.2-3B | 0.613 | 0.773 |
| Pew India | Qwen3-8B | 0.514 | 0.816 |
AUROC predicting whether a simulated answer's error exceeds the median. 5-fold grouped cross-validation. On the US survey the two signals are roughly level; the clear gap is on Indian questions with Qwen3-8B.
What you get
- Local deployment of open-weight models, integrated into the workflows and tools you already use.
- A risk score per answer, read in the same forward pass, to decide what needs a second look.
- Scores that can be logged with every answer, to see where a model fails over time.
- Probes that can be retrained for a new domain without touching the model itself.
What it needs
- Access to the model's weights. A probe cannot be attached to a closed API.
- Labelled examples of that model's mistakes in your domain.
- Validation before use. Extending the approach to agents and tool calls is not yet tested.
Research and production
Our research work is grounded in building systems that have to function outside controlled settings. We have presented our research at top conferences such as ICML (intent recognition from GUI tasks).
We keep up with the research literature as it moves, and we apply a solid understanding of what separates a research prototype from a production-ready AI system, including reliability, evaluation, visibility and control. That lets us help clients bring the latest methods into their operations without the usual gap between demonstration and deployment.
Get in touch
Interested in working together? Tell us about the workflow, data or decision you want to improve, and we'll get back to you.
Contact us