Mohsen Arjmandi

Reasoning, test-time learning, verified autonomy. I build agents that can prove what they did.

Research: ASSAY · ARG · Sensi.

Production: EVON & GRID as CTO @ evolutionID.

Current Work

Agent research and infrastructure: reasoning, test-time learning, evaluation integrity, constrained decoding.

I build AI agents, and I am lucky: the thing I do best is also the thing I love. Fifteen years of shipping production AI, agent runtimes, grammar-constrained decoding, neuro-symbolic agents, distributed systems, and I still work the way I started, passionate and with full attention. I was trained through Iran's exceptional-talents system, and the part I've enjoyed through years has been a simple ask but hard to do: building robots for knowledge workers!

The research work is ASSAY, a general agent harness. The agent is never told what its actions do. It discovers its tools by prediction, declares its own instruments, plans its own subgoals, and every run lands on a hash-chained record anyone can re-verify. First performance benchmark: ARC-AGI-3, RHAE 96.54 under hard action caps. ASSAY is the third architecture in this line. Sensi came first and learned the wrong things. ARG followed and published a null result on arXiv. I kept going, and the third one worked. That is roughly how I do everything.

The production work is at evolutionID, where I am CTO and the architect of EVON, a multi-harness, multi-provider, sandboxed agent runtime in production. The thesis: the session, not the harness, is the canonical compliance and evaluation unit.

Inventor of GRID (Grammar-Railed Decoding, EP 4687028), deployed in production coding agents. Co-founder and CTO of two AI companies. Three granted patents (EPO, USPTO) and two further applications. These works in evolutionID have been winners of the BSFZ Grant for two consecutive years.

Specializations: Reasoning & Agents • Test-Time Learning • Evaluation Integrity • Agent Infrastructure • Access Control for LLMs.

Selected Works

Patents

Granted

Applications

Research

Directed a research roadmap yielding 6 peer-reviewed publications (LREC, SPIE) that directly translated into enterprise product capabilities. Spearheaded the creation of MirasText, the largest automatically generated Persian corpus, to unlock high-accuracy, at-scale Sentiment Analysis, NER, and Topic Modeling for unstructured social streams. Continuously drove platform innovation by deploying Active Learning to slash annotation costs, engineering graph-based algorithms for trend extraction in noisy data, and developing multimodal audio systems (MirasVoice, Acoustic Fingerprinting).

Tools

Applications, tools and platforms I build to help knowledge workers

Writings

Press & Talks