Applied AI research for structured semantic reasoning
Thesis, method, and vision.
Founded by Andres Martinez. Spent 15+ years at Microsoft, most recently as Principal Architect on the VP/CTO team. Multiple patents in enterprise software.
Research Thesis
The reasoning gap
Current AI systems retrieve fragments and generate answers bound to what retrieval returns. The language model has no way to know if the retrieval missed something critical. It reasons over incomplete information without knowing the information is incomplete.
This works for simple questions. It breaks on complex analysis: legal cases spanning hundreds of documents, research synthesis across dozens of papers, enterprise decisions that depend on connecting information across organizational boundaries. In these domains, incomplete retrieval produces incomplete reasoning. Nothing in the system tells you what is missing.
Kacti AI builds the reasoning layer between retrieval and generation. AI systems reason over structure, trace relationships, detect gaps, and report what they found and what they could not find.
Our Approach
Three convictions
These convictions guide the lab's choices.
Structure first, then reason.
Information must be structured semantically before reasoning can be reliable. Connected models of complex information, not embedding indexes, so the reasoning layer works with real relationships, not statistical similarity.
Reasoning lives in systems, not in prompts.
The reasoning layer is a system with state, types, and traceable steps, not a prompt that hopes a model behaves. Durable reasoning needs a substrate: a runtime that holds structure, executes logic deterministically where logic is deterministic, and calls models where judgment is required.
Confidence requires knowing what is missing.
Every analysis must report its own limits: what was found, what could not be found, where contradictions remain. In high-stakes domains, an answer without a stated boundary is a liability.
Structure before code. Meaning before implementation. The lab models a domain's essential concepts and relationships first, then strips everything else away. What reaches the reasoning layer is signal, not noise.
How the lab works
Process discipline applied to AI work itself.
The lab defines explicit processes, guidelines, and guardrails for how its people and AI tools work together. The discipline is inspired by mature engineering practice (CMMI-style phased work: vision, requirements, design, execution) adapted for AI-driven development. The work is repeatable because the process is described, not improvised.
The same discipline ships inside the products.
The Legal Intelligence Assistant works the way the lab works. Pre-loaded workflows, process documents the AI follows, guardrails that prevent drift. The product carries the method, not just the technology.
Semantic modeling as a method.
Before building, the lab models a domain's essential concepts and relationships. Strip noise. Identify the signals that define meaning and behavior. The reasoning layer operates on that model rather than on raw text or statistical similarity.
Method and technology together.
Disciplined process plus semantic modeling is what keeps the lab's systems consistent across domains and what makes the products predictable to use.
Research Areas
Foundational technology
The lab's research spans three areas.
Method
Process discipline for AI development, inspired by mature engineering practice. Semantic modeling as the upfront practice that gives systems their structure. Domain ontologies that translate practice areas into reasoning models.
Algorithms and models
Reasoning systems that operate over structure rather than raw text or statistical similarity. Retrieval becomes a reasoning task; gaps and contradictions surface alongside what was found.
Languages and runtimes
A local graph runtime that holds the connected models the reasoning systems operate on. New language and AI model directions are in progress.
These areas are domain-agnostic at the engine level. Legal is the first applied domain. In legal, the technology runs across complex multi-document reasoning, structured argumentation, and gap detection across large evidence sets.
Foundation first, then domain.
How research becomes product
The lab and the product are the same operation. Research convictions show up in the product as architectural choices: a structured case model, retrieval-as-reasoning, gap reporting.
Legal Intelligence is the first applied surface. It exists because a research thesis about structured semantic reasoning cannot be validated in the abstract. It has to ship and be used by people whose work is measured in case outcomes.
Talk to the founder
The founder is the direct contact for research conversations: collaboration, hiring, investment, deep-dive technical reviews.
The lab works with a small group. Conversations are direct.