Founder & Lead AI Engineer · citronlab.com

Production AI,
end to end.

I run CitronLab — we design, build, and ship production AI for ambitious teams. RAG, temporal memory, on-device LLMs, and the unsexy plumbing that makes any of it work in real systems. Thirty years building software, the last eight in AI.

Visit CitronLab→ Read the white papers

NB 2026

white papers shipped

companies co-founded

30yr

shipping software since 1995

Currently building at

CitronLab

An AI development practice. We take engineering problems where the answer involves an LLM somewhere in the pipeline, and we build the whole thing — model selection, retrieval, memory, evaluation, deployment, the serverless bits.

citronlab.com↗

01 AI engineering, end to end Architecture, build, ship

02 RAG & temporal memory Systems that remember and forget

03 On-device inference Off the cloud, onto the phone

04 Team augmentation EU + US engineers, embedded

White Papers

Six papers on the systems
behind the models.

Original work on consensus pre-processing, temporal memory, two-phase streaming, cold-start serverless, and on-device LLMs. The thinking that informs everything CitronLab ships.

Architecture · Apr 2026

Consensus-Based AI Pre-Processing

Learning on the fly, the fast way

Reduce hallucinations by telling AI the facts instead of asking. A hybrid pre-processing architecture that validates data before LLM reasoning—so AI reasons about implications, not guesses at facts.

09 min read Read paper →

Memory · Mar 2026

Temporal Memory

Why AI systems need to forget

Your AI still thinks you work at your old company. Human memory fades and updates—AI memory should too. A decay architecture that keeps AI knowledge fresh.

12 min read Read paper →

UX · Feb 2026

Two-Phase Streaming

Show results before the model finishes

Users judge speed by when they see something, not when it finishes. Show deterministic results immediately while AI thinks. The streaming architecture that makes slow AI feel fast.

07 min read Read paper →

Performance · Jan 2026

Cold Start Zero

Rust for real-time serverless

That 500ms delay your users feel? It's Python's cold start. Rust drops it to 10ms. Same code, 10x faster, 10x cheaper. Here's how to make the switch.

11 min read Read paper →

Privacy · Dec 2025

On-Device LLM

The end of cloud AI dependency

Apple Silicon + MLX = real AI on your laptop. No API calls, no privacy concerns, no monthly bills. Works offline. Here's how to set it up.

10 min read Read paper →

UX · Nov 2025

Proactive AI

Surfacing what you need before you ask

Before your meeting with John, AI surfaces: last conversation, open commitments, relevant docs. Stop asking—start receiving. The architecture behind anticipatory AI.

08 min read Read paper →

Photography