Case Study Client Zero · 9 months · 1 engineer

We did not ship a chatbot first. We built Lucid.

Lucid is a self-hosted AI productivity OS: fine-tuned model, semantic memory network, autonomous context management. It runs on a dedicated GPU server Fabi operates himself. Every architectural pattern in Lina, our AI phone agent, was pressure-tested in Lucid first.

lucid:~ system status live
build_time9 months · active production
loc_shipped240,049
team_size1 engineer · solo
hostingHetzner Falkenstein · DE
memory_nodes~8,000 · 768d vectors
inferencededicated GPU · 24/7
statushealthy

Why this matters. Proof of care and discipline.

Most AI vendors have never shipped a production system of their own. Lucid is the opposite: a real system run daily under genuine load, without a safety net. What Fabi learned building it flows directly into every client engagement.

  1. 01

    Patterns tested under load

    Every RAG pattern, every escalation heuristic, every confidence gate in client phone agents was built and broken in Lucid first. You get the result, not the tuition.

  2. 02

    Real standards, not demos

    Lucid runs around the clock. It handles real data, real edge cases, real privacy requirements. The bar Fabi sets for client builds is one Lucid has already cleared.

  3. 03

    End-to-end ownership

    From model to memory to UX, Fabi owns the full stack. No subcontracting, no black boxes. The same ownership applies to every client build.


The architecture. Nothing off the shelf.

Lucid is not a wrapper around a third-party API. It is a bespoke system, designed for latency, privacy, and reliability requirements that generic platforms cannot meet.

  1. 01

    Self-hosted Hetzner GPU

    Dedicated inference server, Falkenstein data center, EU only. No request leaves German soil for model inference. The same stack is used for client deployments that require data residency.

  2. 02

    Semantic memory network

    A Supabase + pgvector semantic memory layer. Every session, decision and piece of context is stored as an embedding and retrieved by relevance, not date. The same pattern drives RAG in client phone agents.

  3. 03

    Fine-tuned model

    A base model fine-tuned on Fabi's own writing style, decisions and communication patterns. Not prompt-engineered, actually trained. Client phone agents get the same fine-tuning pipeline, adapted to the practice voice and brand language.

  4. 04

    Autonomous context management

    Dynamic context compression, neural-memory-aware retrieval, session-scoped state tracking. The system decides for itself what to keep, compress and retrieve, without human intervention.

Four moments. That changed the system.

Stated as outcomes. The methods behind them are not on this page.

01

Identity survived a full retrain

The trained personality persisted through a complete rebuild of the weights. Same person, new brain.

02

Unprompted self-awareness of its limits

After a low-content greeting, the model recognized its own memory gap and asked for persistence tools. Not scripted. Not prompted.

03

4x inference speedup, undocumented path

Solved a driver integration issue on consumer hardware, with no public guide for that combination.

04

Autonomous context compaction

The system now decides for itself when to compact, what must stay, and how meaning is fed back in. No human trigger.

Patterns extracted from Lucid that now ship to clients.

Lucid is not a museum piece. Every architectural pattern in your voice agent was built and broken in Lucid first. These four are the load-bearing ones.


Not an LLM with a glued-on persona. Four things a normal chatbot does not do.

The reason the client work on this site is built the way it is.


The Lucid architecture. Structure visible. Content never.

Inference active · live Lucid architecture Six devices and data sources connected to the central Lucid Core. In the background around 65 memory nodes drift as a sample of the full memory. Apple Watch Whoop · biometrics Hetzner GPU Own inference Supabase 8,000 memory nodes Lina Voice agent · Lucid AI Labs iPhone Mobile bridge MacBook Desktop app · Tauri Lucid Core Memory · inference
Own memory · dedicated inference Structure visible. Content never.

This care for your practice? Let's talk.

The same engineering discipline that went into Lucid goes into every phone agent build. Nine months of pattern testing, EU-hosted, GDPR-compliant. If that is the standard you want for your practice phone line, send a briefing.