We build systems that stay close to the world they serve. Human-first. Rigorously empirical. Unapologetically practical. Across three research pillars, we investigate where AI actually breaks - agents that fail on contradictory enterprise data, benchmarks that miss what matters in real classrooms, models too expensive for the people who need them most.
“A model that only performs under ideal conditions has not solved the problem. It has avoided it.”
– The Philosophy
Read our philosophy →Publish research that is instantly actionable for practitioners – not locked behind paywalls or buried in jargon.
Build AI systems that work on affordable hardware, not just on cloud clusters that cost more than the schools we serve.
Develop open datasets and benchmarks rooted in lived community knowledge – measuring what matters, not what is convenient.
Equip students with the tools to lead their own inquiries, not just consume ours.
Papers presented at international conferences
Team grows to 23. Eleven new projects launched
Three provisional patents filed
First papers submitted to peer review
Research briefs take shape. Three pillars defined
First LLM experiments. The idea of a lab germinates
Solve real problems and guide researchers on weekends while building your applied research muscle.
Source grounded, publishable challenges for your UG, MS, or PhD students to work on.
Tackle field-tested problems with direct guidance from both industry and academia.
Sponsor or suggest research directions based on the bottlenecks your company actually faces.
Bring your applied experience and contribute to open, grounded intelligence.
Three pillars. Fourteen projects. One mission: build intelligence that works in the real world.
Failed coding agent runs contain real diagnostic work that gets thrown away. We build recovery agents that read those traces and finish the job at a fraction of the cost.
Enterprise AI agents act on contradictory data across systems and never notice. We build agents that traverse Slack, CRM, and contracts to find and reconcile those conflicts.
Security logs contain evidence of attacks that no pre-built rule anticipated. We investigate whether an AI system can review yesterday's logs and surface attack chains on its own.
Document AI models extract the right values but assign them to the wrong fields, and no benchmark measures this. We build an evaluation that tests concept-binding across paystubs, invoices, contracts, and tax forms.
Vibe coding tools produce apps that work in demos but fail on mobile, security, and accessibility. We build a benchmark that scores the running app in the browser, not the generated code.
AI memory systems benchmark on recall, not on whether they actually change behavior for each user. We build an evaluation that tests whether systems adapt tone, timing, and restraint across sessions.
Handwriting recognition treats every page as if it came from a stranger, even for writers it has seen before. We investigate persistent writer adaptation that learns once and recognizes at full speed from then on.
Multilingual AI requires expensive retraining to reason in each new language. We investigate whether reasoning and language grounding are separable, so one core model serves many languages by swapping only embeddings.
Training an AI tutor by rewriting the entire model is expensive and risks degrading the reasoning it teaches. We explore architectures that separate teaching strategy from subject knowledge, so pedagogy trains cheaply and ports across models.