Michael Fester is the co-founder of GloGlo and 14.ai, companies applying artificial intelligence to consumer brands and customer-support operations. He previously co-founded Snips, whose privacy-focused, on-device voice technology was acquired by Sonos in 2019.
Fester left doctoral research in number theory at Cambridge to build companies and created 8pen, a handwriting-inspired mobile text-entry system. At Snips, he helped develop voice assistants that processed requests directly on connected devices, reducing dependence on remote servers. He later co-founded Motif, a collaborative technical-documentation platform, followed by Markprompt, which connected company knowledge to AI-powered documentation and support through content ingestion, developer APIs, embedded interfaces, analytics, and vector storage. Its analytics helped teams identify unanswered customer questions and improve their documentation.
Fester and co-founder Marie Schneegans joined Y Combinator’s Winter 2024 batch with Markprompt. Their work expanded into 14.ai, combining purpose-built software, AI agents, and human judgment in a managed customer-support operation, and GloGlo, a glucose-focused consumer brand using their AI infrastructure.
Building support systems that remain dependable
- Deterministic guardrails for autonomous support: Fester organizes agents around focused actions, explicit workflows, and domain-specific sub-agents. Canceling a subscription becomes a controlled sequence of eligibility checks, retention decisions, and account changes, with branching, retries, state transitions, and memory represented explicitly.
- Typed, observable agent infrastructure: His Effect and TypeScript architecture shares schemas across React interfaces, RPC services, public APIs, generated documentation, and Postgres. It supports model-provider failover, stateful retries, duplicated token streams, OpenTelemetry instrumentation, and dependency-injected failure simulations.
- Practical adoption over architectural absolutism: Fester recommends introducing Effect incrementally while acknowledging its learning curve, the difficulty of tracing dependencies in larger systems, and the danger of error handling that silently conceals meaningful failures.
