Verification
Independent proof that a result is actually correct.
sagtuku. An AI + X neo lab by Alpai GmbH.
Intelligence
That Learns
From Reality
We build AI that takes on real goals, does the work, checks the results, and learns from what actually happened. Not from what it expected.
- Steps in the loop
- 07
- Research programs
- 09
- Research agent
- PH.I
01The missing piece
AI can do the work. It can’t yet tell if the work is right.
Today’s best AI can reason, write code, and use tools. But to work on its own, it still lacks five basic skills.
“The next step is teaching machines consequences.”
- 01Verify
Tell whether an action really worked.
- 02Persist
Stay on one goal for days, weeks, or months.
- 03Learn
Get better from real experience.
- 04Predict
Foresee what an action will cause.
- 05Transfer
Reuse skills in new settings, from code to machines.
02The closed loop
AI that gets better by doing.
Most AI is trained once and then stays the same. Our systems run in a loop. Every round adds experience and evidence to the next.
FIG. 01THE CLOSED LOOP
01Reason02Plan03Act04Observe05Verify06Remember07Learn
- 01Reason
Understand the goal, the limits, and what success looks like.
- 02Act
Do the work with approved tools and code.
- 03Observe
Look at what really changed. Don't assume the plan worked.
- 04Verify
Check the result with tests the agent does not control.
- 05Remember
Record what happened, including failures and sources.
- 06Learn
Turn checked experience into better skills.
- 07Reason better
Start the next task knowing more than the last.
03What we research
Everything around the model.
Models are getting smarter. We work on what they still need to finish real tasks: staying on track, checking results, and learning from mistakes.
Long-running agents
Agents that stay on track through long, complex tasks.
Memory & learning
Turn successes and failures into lessons for future work.
World models
Predict what actions will cause, and carry skills into new settings.
4 of our 9 research programsSee the research
04Phase I · Research agent
Give it a goal,
not a prompt.
Our first product is an AI agent for technical research and engineering. We start with digital work because every step can be inspected and repeated.
Find out if architecture X can cut inference costs without hurting performance. Design and run the experiments, analyze the results, question your own conclusions, and show your work so others can repeat it.
Goal received · Proof required
What you get back
- 01Hypotheses
- 02Code
- 03How each test was run
- 04Raw results
- 05Reasoning steps
- 06Checks passed
- 07Confidence level
- 08Evidence against
- 09Steps to reproduce
Software engineeringAI researchData analysisExperiment designModel evaluationComputational research
See how a task runs05How we measure success
Not test scores. Results that hold up.
Time on task
How long can the AI work on a real goal before a person has to step in?
Verified results
How many finished tasks pass an independent check?
06Questions
Straight answers.
01Who is behind sagtuku?
sagtuku is an AI + X neo lab operated by Alpai GmbH, a Swiss AI company. "AI + X" means we pair AI with a real field of work (X), such as software, science, or engineering, and build systems that deliver results in that field.
02What is closed-loop intelligence?
It is AI that does not stop at an answer. It acts, looks at what happened, checks the result, and remembers the lesson. Each task makes it better at the next one.
03What is sagtuku working on first?
An AI agent for technical work: software engineering, AI research, data analysis, experiment design, and model evaluation. We start here because every step and result in these fields can be checked and repeated.
04How does the agent know its work is right?
It does not take its own word for it. Separate checks test the result: automated tests, simulations, critical reviews, statistics, or repeated experiments. When the evidence is unclear, a person makes the call.
05Is sagtuku building another chatbot or a bigger model?
No. We build the system around the model: long-running tasks, independent checks, memory, and learning from experience. The goal is AI you can trust to finish real work, not a new chat window.
↗Work with us
Give us a goal, not a prompt.
Working on a hard research or engineering question? Start with the result you need to prove.