Engineering Behavior Observatory
Working toolCaptures AI engineering-agent trajectories and workspace outcomes for evidence-grounded evaluation and comparison.
TypeScriptNode.jsOpenTelemetry
Applied AI research and development.
Kumanday Intelligence is an applied AI lab. We study how agents, models and generative media behave on real tasks, and we turn the results that hold up into working tools for our clients.
Mission and vision
We are a research lab first. Our engineering exists to carry research results into systems people use every day.
Mission
Run applied AI research that ends in working systems, with results our clients can measure and the community can reuse.
Vision
A future where people and AI systems work together on hard problems, and every claim about what AI can do is backed by evidence.
Research lines
Six lines of work, each with a question we are trying to answer and, where the evidence is in, a tool that carries the answer.
Agent systems
How coding and computer-use agents plan, coordinate and recover. Parallel coding agents, multi-agent protocols such as A2A and MCP, and desktop automation.
Active OpenSymphonyEvaluation and observability
Measuring what agents actually did. Trajectory capture, reference-free LLM evaluation with G-Eval, and benchmarks for multi-agent coordination.
Active Engineering Behavior ObservatoryGenerative media
Video generation, narration and automated editing, from keyframes and chained segments to verified final renders.
Active TTV Pipeline, media-toolingAI for learning
Adaptive tutoring, personalized lesson paths and AI-generated teaching visuals, tested with students and teachers.
Active Parlina lesson playerModels and inference
Distillation for local inference, long-context serving, and newer architectures including diffusion language models, Mamba and Titans.
Exploring Notes and experimentsReal-time multimodal
Streaming audio and video interaction, ambient agents, and sandboxed environments where agents can use a computer safely.
Exploring Notes and experimentsTools in use
Software that came out of the research lines and now runs in client work. Each one is open for inspection.
OpenSymphony
Working toolCoordinates coding agents through Linear issues, isolated workspaces, dependency scheduling and human review.
RustLinearOpenHandsACP
media-tooling
Working toolTurns source media into transcripts, subtitles, planning artifacts, rough cuts, narration and verified renders.
PythonFFmpegHyperFrames
TTV Pipeline
Working toolPlans and generates video segments through keyframe and chaining workflows, with approved generation records for editorial handoff.
PythonFFmpegRedisRQ
Earlier experiments
cli_engineer, an agentic CLI in Rust with more than 1,000 downloads in its first five days · a GraphRAG proof of concept · a reference-free LLM evaluation framework · Nexus Agents, a hierarchical multi-agent research system.
Clients
We work as an embedded research partner. Our clients bring the problem and the setting; we bring the experiments, the evidence and the tools.
Trilogy
Research and tooling for how engineering agents work: what they do, how to run many of them at once, and how to produce video with them.
- Lines
- Agent systems, evaluation, generative media
- Tools
- Engineering Behavior Observatory, OpenSymphony, media-tooling, TTV Pipeline
Parlina
An AI learning platform for universities and training programs. We research and build adaptive lessons, the AI lesson player and tutoring workflows.
- Lines
- AI for learning, generative media
- Work
- Lesson player, adaptive content, AI tutoring
Teams we have worked with
PayPal / Cisco / eBay / Yahoo / Nokia / McKesson / IHME / Velolegal / Safekeeping
How we work with clients
Every engagement starts with a question. What we deliver depends on how far the answer needs to travel.
Research sprint
A defined question, a few weeks of experiments and a written result you can act on.
Evaluation
We measure agents and models on your own tasks before you commit to one.
Prototype to tool
Working software that carries a result into daily use, with the evidence attached.
Embedded R&D
A standing research function alongside your team, with regular readouts.
Podcast
Where we talk through the research in public.
Synthetic Intelligence
Episodes on AI research, autonomous agents and what it takes to build with them, from new papers to the tools we use in our own work.
AgentsEvaluationGenerative mediaResearch papers
Listen on YouTube ↗Community and internships
We teach what we learn, and we train the next researchers by putting them on live work.
Masterclass: multi-agent architecture for deep research and generalist agents
MasterclassTake the quiz →AI Tinkerers Manizales, fifth meetup
MeetupResearch internships
Interns join a live research line and work with mentors on real client problems.
- Academic credit with partner universities in Colombia and Latin America
- Thesis projects in real settings
- Support for papers and conference talks
- A path to a full-time research role
Team
Experience in software and data, applied to research.
Leonardo González
Founder and CTO
Leonardo has more than twenty years of leadership in software engineering, architecture and data science. He has led teams on multimillion-dollar projects, advised startups as a consulting CTO, and held architecture and technical leadership roles at Fortune 500 companies.
He leads Kumanday's research on agents, evaluation and generative media.
Start with a question you want answered.
Tell us the problem and the setting. We will reply with how we would test it.