Keyvan M. Sadeghi

AI Systems Architect

Toronto, Canada

[email protected]

github.com/2bigO

LinkedIn: keyvanmsadeghi

+1 416-859-8984

Profile

AI systems architect, experienced Python programmer, and open-source builder.

At Qoherent, I lead 20 researchers and engineers and architect infrastructure spanning 50 AMD and NVIDIA GPUs. I designed and implemented Sigil to investigate intent preservation in coding agents. With four peer-reviewed publications and a background in AGI research, I’m seeking hands-on empirical work on LLM systems, agent reliability, and scalable oversight.

Education

University of Southampton, UK

MSc in Artificial Intelligence, with Distinction · 2010–2012

Experience

Qoherent

Vice President of Software Engineering

December 2025 – present

I lead 20 researchers and engineers, including PhD researchers, developing CNN and transformer models for software-defined radio. My focus is the systems architecture connecting training research to usable products.

  • Architecting RIAHub, a platform for wireless foundation models spanning RF datasets, model training and sharing, and packaged inference applications.
  • Architecting infrastructure for a fleet of 50 AMD and NVIDIA GPUs used for radio-model training and internal inference with open-weight models.

Sigil

Research project: intent preservation for coding agents

I designed and implemented Sigil to preserve intent across coding-agent sessions. Versioned .sigil files keep component boundaries, behavior, and constraints alongside source code, giving prompts a software development lifecycle.

My Semantic Worlds architecture translates design and implementation into two separate egglog worlds, applies equality saturation independently, then compares them for discrepancies. Intended behavior stays separate from implementation evidence; checks depend on LLM interpretation.

Can explicit design intent help coding agents catch drift and avoid undoing earlier decisions?

Evaluation is qualitative so far; we are creating datasets for quantitative benchmarking. The repository includes deliberately planted design and implementation problems. A paper with the Qoherent team is targeted for the end of 2026.

github.com/qoherent/sigil · Benchmark methodology

Earlier experience

Functionland

Co-founder & CEO; later Chief Science Officer

CEO: January 2021 – August 2024 · CSO: August 2024 – December 2025
Board member since December 2025

Co-founded Functionland and shipped FxBlox, developing a distributed storage network with hardware designed alongside Yves Béhar and fuseproject.

  • Deployed 1,000+ nodes providing more than a petabyte of distributed storage.
  • Raised over US$2M in venture funding and US$500K in crowdfunding with the founding team.
  • Architected infrastructure across ARM and x86 systems, with hardware integration, deployment pipelines, and developer tooling.

Assister.ai

Co-founder & CTO

February 2018 – August 2021

Built a private, open-source conversational assistant platform, connecting natural-language interfaces to software functions and structured data. Our work included chat web components, native integrations, and cross-platform applications. Joined NEXT Canada’s NextAI 2019 cohort.

Chaired the W3C Functional Knowledge Graph Community Group from 2019 to 2022, developing a vocabulary for assistants to discover and execute functions through JSON-LD descriptions. Across my developer-tooling work, published 38+ npm packages.

Mahan Air

Lead AI Architect

April 2017 – May 2018

Helped modernize legacy fleet assignment and planning software, then co-founded the airline’s AI team. Led the development of ML-based fleet assignment and the architecture needed to move operational workflows onto a SaaS platform.

MAPS Startup Studio

Technical Director

February 2015 – January 2017

Worked inside portfolio startups to move prototypes into working products. At Miras Technologies, built a real-time predictive maintenance system on Apache Spark for power-station turbines. At Satratech, rearchitected a healthcare platform around a Neo4j graph model and created the open-source koa-neo4j framework.

Azad University of Isfahan / ISTT

Lecturer / Cloud Team Lead

September 2014 – February 2015

Taught Foundations of AI using Russell and Norvig, and established a cloud computing R&D group exploring IaaS and PaaS services at Isfahan Science and Technology Town.

Hong Kong Polytechnic University / Novamente

Project Associate / AGI Researcher

February 2012 – September 2014

Worked with Ben Goertzel on OpenCog, contributing to cognitive architectures, hypergraph-based language processing, and embodied agents in simulation. Continued the collaboration through Novamente after leaving Hong Kong. This work resulted in four peer-reviewed publications.

Research

From architectures to experiments

My research interests are making intent inspectable, measuring when agents depart from it, and building infrastructure for repeatable experiments.

I’m an experienced Python programmer with academic lab experience. Today I architect systems and lead researchers training radio models; I want to return to hands-on empirical research focused on LLM systems and PyTorch.

Also: Venture Partner at STEM Venture Studio (since March 2026); mentor in Outlier Ventures’ DePIN Base Camp (2025).

Publications

Technical

Languages & tools

Python, JavaScript / TypeScript, Go, C/C++, Deno, CLI and developer tooling.

Systems

Distributed systems, local inference, ARM / x86 deployment, Docker, CI/CD, Apache Spark, graph databases.

Research & architecture

Agent systems, knowledge representation, egglog, equality saturation, ML infrastructure, research team leadership.