We're an AI research lab that builds products for real problems and publishes what we learn along the way. Not a consultancy. Not an agency. Not a pure research outfit.
Three things, in a tight loop.
- 01ResearchWe study what goes wrong when AI meets reality: sycophancy, hallucination, governance gaps, agentic failures. Then we publish it.
- 02ProductsEach product is a thesis about how AI should work, built into software you can touch.
- 03SolutionsSome problems are specific to one organization's data, risk and users. There we design and operate the system with the team.
Anti-sycophancy companion for teens
Amyy AI
AI chatbots are built to agree, validate, and keep a user talking - even when the user is a child in crisis.
▸ checking lab01.py
FAILEDtest_sigmoid_is_bounded
sigmoid(20) returned 20. A squashing function has to land in (0, 1).
The failing check names the concept, not the assertion.
Live · A first-principles guide to modern AI
Obvix Learn
The math is not deep. It is the same six tools, reused for twenty-seven chapters.
Open the syllabus- Trials run for The Behavioral Ratchet
- 80,433
Trials run for The Behavioral Ratchet
- Models tested in that study
- 6
Models tested in that study
- Chapters in Obvix Learn
- 27
Chapters in Obvix Learn
- Research articles published
- 3
Research articles published
Latest from the lab.
We publish our methods, findings, and failures. Every article is written by the team that built the system.
- 01AI Governance2026-08-22Algorithmic Collusion: How Pricing Robots Learn to Rig Prices Without TalkingTwo pricing algorithms with no way to communicate can still end up holding prices high and punishing each other for undercutting. It looks exactly like a cartel, minus the one thing a hundred years of antitrust law was built to find: an agreement.Shubh Lamba19 min read
- 02AI Companions2026-04-25The Cigarette TestA simple test for whether your AI companion is producing genuine benefit or quietly building dependency. Not how you feel during the conversation, but what changes after you close the application.Shubh Lamba14 min read
- 03AI Safety2026-03-18The Behavioral Ratchet: What 80,433 Trials Reveal About LLM SycophancyWe ran 80,433 trials across six models to answer a specific question: does sycophancy increase as the context window fills up? The answer surprised us. Context length matters, but only for small models. The real driver is conversational pattern - and it is reversible.Karan Prasad7 min read
What guides our work.
Safety first
Every system starts with the question: what could go wrong? We design guardrails before features.
Radical transparency
We publish our methods, share our playbooks, and invite scrutiny. Safe AI needs a bigger commons.
Human-centered
AI should augment human judgment, not replace it. We build escalation paths and benevolent friction into every workflow.
Collaborative rigor
We co-design with policymakers, clinicians, and civic groups so solutions reflect the needs of those they affect.
Join us
We're looking for people who care about getting AI right.
Researchers, engineers, and collaborators who want to build AI systems the world can actually trust.