01
Research
Understanding real-world behaviour
We study where general-purpose models lose context, mishandle sources or fail to represent the people and knowledge environments they serve.
Research / products / safety
Collaborative
Intelligence
Research-born. Context-aware.
Built to be tested.
What we do
We build AI systems with the people and contexts they are meant to serve.
Our work combines applied research, product engineering, domain expertise and evaluation. We focus on the gap between general model capability and reliable performance inside real communities, cultures and high-trust domains.
Our approach
Research grounded
in real contexts.
We are exploring how AI systems can be made more useful, culturally aware and accountable in the communities they affect.
02
Evaluation
Testing beyond aggregate scores
We are developing approaches to domain-specific evaluation, adversarial testing, expert review and qualitative failure analysis.
03
Prototyping
Learning through bounded products
We build focused prototypes to test research questions in practical settings before making broader claims or commitments.
Applied AI research
Evaluation.Adversarial testing.Benchmarking.Domain alignment.
We work on behaviours that disappear inside aggregate benchmark scores: cultural context, source fidelity, disagreement, uncertainty, multilingual performance and failure under domain-specific pressure.
Capabilities
Areas we are
exploring.
Domain-specific evaluations
How task suites, rubrics and expert-labelled cases can reveal model behaviour inside specialised knowledge environments.
Adversarial and red-team testing
Targeted ways to examine hallucination, source misuse, unsafe deference, cultural failure modes and brittle reasoning.
Benchmark design
Representative datasets and evaluation protocols that could measure more than generic accuracy or preference.
Expert feedback loops
Ways to capture review, disagreement, correction and provenance without flattening expertise into a single score.
Grounded AI products
How retrieval, citations, structured context and human review can support more grounded product experiences.
Safety-oriented pilots
How bounded pilots can generate evidence about model behaviour before capability is expanded.
Research questions
General capability is not the same as local reliability.
How should models behave when sources conflict? How do we evaluate culturally situated judgement? What does calibrated uncertainty look like in a user-facing system? How can community expertise improve testing without becoming invisible data labour?
These questions shape our products, evaluation work and research partnerships.
Our approach
Build, test, learn, repeat.
Our approach combines prototype observation, structured human review, qualitative failure analysis and adversarial evaluation. The objective is to produce evidence about where systems work, where they fail and what changes improve them.
We are early, independent and selective about what we claim. Our first results are published: AICAB asked ten models fifty questions on Muslim life, belief and culture, three times each, and every one of the 1,500 answers is open to read.
Community knowledge should shape how AI systems are trained, tested and evaluated—not only how they are marketed.
Partnerships
Work with us.
Model and research labs
Exploratory evaluations, culturally specific red teaming, benchmark research and analysis of domain-level failure modes.
Scholars and domain institutions
Expert review, evaluation design, corpus governance and carefully bounded research pilots.
Safety and evaluation teams
Adversarial test design, multilingual and cross-cultural coverage, judge calibration and qualitative error analysis.
Compute and infrastructure partners
Inference, experimentation, secure data environments and infrastructure for responsible product-scale testing.
Funders and mission-aligned investors
Patient capital and research support for infrastructure that communities can trust and meaningfully shape.
Start a conversation
Building AI with communities, not merely for them.
If you work on frontier models, evaluations, AI safety, domain expertise or mission-aligned infrastructure, we would like to hear from you.