Alessio Fanelli is an engineer, startup investor, and co-host of Latent Space, the AI-engineering podcast he founded with Shawn “swyx” Wang. A partner at Decibel who is also building Kernel Labs, he backs technical founders, interviews leading AI practitioners, and develops infrastructure for coding agents.
From open hardware to AI engineering
Fanelli grew up in Rome, taught himself programming, and studied physics and computer science before leaving university to build Smart Torvy, an open-source, open-hardware home-automation startup inspired by a campus hackathon. He and his collaborators exhibited their technology at Maker Faire.
After engineering roles at early-stage companies including Welcome Tech, he moved to the United States for an engineering fellowship at 645 Ventures. There, he built software for identifying and analyzing startups, became a principal, and backed developer-tool, security, and enterprise-software companies including Cube Dev, Panther Labs, Launchable, Oort, and Bigeye. He also contributed to Cube Dev’s open-source software. At Decibel, he combines early-stage investing with internal software development and advises founders on product strategy, recruiting, and go-to-market execution.
Fanelli and Wang launched Latent Space in 2023 to give engineers building AI products technically substantive conversations about models, infrastructure, developer tools, and applications. Fanelli describes its editorial approach as balancing durable technical explanations with timely conversations around consequential releases, supported by written summaries, transcripts, and chapters.
He automated parts of that publishing workflow through smol-podcaster, an open-source project that produces speaker-labeled transcripts, chapters, suggested episode titles, and promotional copy.
His distinctive ideas and projects
Startups are context arbitrages. As coding agents reduce the cost of implementation, Fanelli argues that defensible companies gain their advantage from specialized knowledge and operational context: understanding threats, customer needs, or infrastructure problems better than their competitors.
Shared infrastructure for coding agents.Kernel Labs develops tools including AWT, a coding-agent worktree manager; Glimpse, which captures screenshots during continuous integration; and Kernel Gym, for testing and sharing Model Context Protocol servers.
Agent evaluation must reflect real behavior. In experiments with GPT-5 and Claude Opus, Fanelli found that models tasked with building developer utilities largely ignored those utilities during subsequent application-modernization work. His continuously running coding-agent setup consequently emphasizes isolated workspaces, issue tracking, review gates, merge queues, and per-task token accounting.
At AI Engineer World’s Fair 2025, Fanelli interviewed Sierra cofounder Clay Bavor about enterprise AI architects, customer-experience teams, agent coaching, and build-versus-buy decisions. Sierra’s products and customer examples belong to Bavor’s company; Fanelli’s contribution was probing the organizational and engineering demands of deploying customer-facing agents.
