Devlar Technologies
Applied AI research, tested in production.
Devlar is an AI research company. We study why autonomous agents fail, how AI systems should remember, how to spend compute only where it matters, and what makes people trust a personal AI. Every idea is tested against a working system. Founded by a former Adobe Senior Engineering Manager with 9 years building production systems.
Research areas
Four open questions about AI systems that have to work outside the lab.
Agent reliability
Why do autonomous agents fail in production, and whose fault was it?
Failure modes of multi-step agents, telling a broken credential from a broken upstream from broken logic, and judging a run by what it produced rather than how it exited.
Projects: AI Workforce, Docket, Agent Starter Kit
Memory & context
How should an AI system remember?
Persistent, structured and editable memory that carries context across sessions without dumping transcripts into the prompt.
Token efficiency
How little frontier-model compute does a task actually need?
Agent loops quietly multiply token spend. We study layered architectures that reserve large models for reasoning nothing cheaper can do.
Projects: AI Workforce
Human-AI interaction
What makes a person trust a personal AI?
Privacy, ownership of your own data and memory, and how people build working relationships with an AI companion over time.
Projects: Alvi
Why Research
After 9 years at Adobe as a Senior Engineering Manager, including leading the team that built a production AI agent for real-time Slack incident summarization and Jira automation, our founder kept meeting the same problems: agents that fail silently, forget what they knew, and spend far more compute than the task needs. Those are not tooling gaps. They are open research questions, and Devlar exists to work on them.
Research projects
We test every idea against a working system. These are the systems.
AI Workforce
In progressAgent reliability · Token efficiency
Can a team of specialist agents carry a multi-step goal end to end, reliably and at a predictable cost?
Docket
In progressAgent reliability
Automatic fault attribution: was it the credential, the upstream API, or the workflow's own logic?
Agent Starter Kit
Reference buildAgent reliability
What is the smallest agent scaffold that is ready for production rather than ready to demo?
Synapse
In progressMemory & context
Structured, editable, consented memory for AI systems, in place of transcript dumps.
Alvi
Early stageHuman-AI interaction · Memory & context
A personal AI companion owned by the person using it, built to study privacy, ownership and trust.
Our projects are instruments, not products: each one exists to answer a research question under real workloads
Why Devlar
The AI agent space is crowded with demos. We study what happens after the demo: why agents fail, how they should remember, and what they really cost to run.
Production Is the Lab
Every question is tested against a working system under real workloads, not only against a benchmark. If an idea cannot survive production, we want to find that out.
Measured, Not Demoed
Every agent action, tool call and token is logged as a structured event, so a claimed improvement can be traced to the change that made it.
Grounded in Real Problems
Our research agenda comes from direct experience shipping AI agents in production. The questions are ones our founder has actually had to answer.
Grounded in Production Experience
Devlar's research questions come from an engineering leader who has shipped AI into production, not as a prototype but as a tool used daily inside one of the world's largest software companies.
9 Years at Adobe
Founder Alan O'Meara spent 9 years as a Senior Engineering Manager at Adobe — leading infrastructure engineering, large-scale platform migrations, and the development of internal AI tooling used across engineering teams.
Led a Production AI Agent Team
Alan led the team that built a production AI agent deployed inside Adobe that automated Slack war room summarization and Jira ticket updates during live incidents, reducing manual overhead in high-pressure engineering situations.
From Production to Research
Running AI in production surfaces problems that benchmarks miss: silent failures, lost context, runaway cost. Devlar turns those into research questions and tests the answers on working systems.
Follow the research
Four research areas, five working systems. See what we are investigating and where each project stands.
See the research →Working on the same problems?
We want to hear from researchers, engineers and teams who have hit the same walls: agents that fail silently, forget, or cost more than they should. Collaborators and investors are welcome too.