Work Production systems, not demo projects

What we built. Under real load.

One flagship project: Lucid. Nine months of building, breaking and rebuilding a production AI system under daily live load. Every pattern in our voice agent builds was pressure-tested here first.


Case study 01

Lucid. Personal AI system, production-ready.

Start: September 2025. One developer. No co-founder, no team, no external funding. A full-stack AI productivity companion built to the same standard we apply to every client build, because it was the testing ground for that standard.

9 mo. active build
240k lines of code
1 developer
Daily production load

The stack that runs it

Inference
Hetzner GPU · Falkenstein, DE
Memory
Supabase + pgvector
Model
Fine-tuned · custom RLHF
Client
iOS · React Native
Context mgmt.
Autonomous · custom hooks
Orchestration
n8n · custom tool chain
Full Lucid deep dive

Case study 02

The Safeguards Architecture. Engineering pattern, in every build.

Every Lina deployment ships with four safeguards baked into the request pipeline. Not as a configuration option, not as a premium tier. Architectural defaults that pass every call through identity disclosure, a confidence gate, scope enforcement and full transparency logging, in that exact order.

< 2s warm handoff
0.75 confidence threshold
100% transcripts logged
First sentence: identity
  1. 01

    Identity disclosed, first sentence

    The bot announces itself in the first sentence of every conversation. "Hi, you are speaking with Lina, the digital assistant from [Practice]." Not in the small print, not buried, not optional. The pattern is architectural, not policy.

  2. 02

    Confidence gate at 0.75

    Every model response is scored against a calibrated confidence threshold. Below 0.75: clean decline + handoff path, not hedged speculation. The bot never guesses on medical-adjacent questions because the model cannot bypass the gate.

  3. 03

    Scope check, then clean decline

    Scope is defined explicitly at build time per practice. Out-of-scope requests trigger a clear decline ("That is outside what I can help with, let me transfer you") rather than confidently-wrong hedging. Endless hold-loop patterns are architecturally ruled out.

  4. 04

    Full transcript logged, real time

    Every conversation turn writes to the admin dashboard with timestamps, confidence scores, and handoff reasons. You and the caller can request the log at any time. No black-box decisions, no plausible deniability either way.

Read the safeguards deep-dive

Methodology · how case studies happen

Every build passes 6 gates.

Case studies are not written first and built after. They are the output of a process that runs before a single line of production code ships. The gates below are non-negotiable for every Lina deployment, in this exact order.

  1. 01

    Scoping intake

    One structured intake brief, by email, before any contract. We audit your current phone setup, call volume and booking system in writing. If the fit is wrong, we say so here, not after the build starts. No scheduled phone calls in this process.

  2. 02

    Brief

    A written scope document covering voice persona, escalation rules, booking integrations and explicit out-of-scope topics. You review and sign before build begins. No ambiguity enters the codebase.

  3. 03

    Build

    RAG knowledge base ingestion, safeguard layer wiring, booking-system integration and voice persona configuration. Each is tested in isolation before the pipeline is assembled. Average build window: 10 to 14 days.

  4. 04

    Pressure test

    A structured adversarial call session: off-topic probes, confidence-floor triggers, handoff edge cases. We run these before you hear the voice. Failure here delays launch; failure at a live caller does not happen.

  5. 05

    Soft launch

    Two weeks of parallel routing: a portion of real calls go to Lina while your existing line stays active. Transcript review with you every 48 hours. Full cutover only after both sides are satisfied with the numbers.

  6. 06

    Operate

    Monthly knowledge-base updates, confidence-threshold tuning as call patterns shift and admin-dashboard access so you see every conversation without needing to ask. The case study gets written here, once the numbers are real.


Case study 03 · 04 · 05

Three founding seats. No invented testimonials.

The voice-agent product is new. The architecture isn't, Lucid above is the proof. As each founding clinic goes live, they appear here with deployment details and outcomes. No padded logo grid, no faked quotes. This is the current state, in writing.

  1. 03

    Open, first founding seat

    €449/month in year 1, €0 setup. Case study posts here when this clinic is live.

  2. 04

    Open, second founding seat

    Same terms. Reserved as soon as the first slot signs.

  3. 05

    Open, third and final founding seat

    Last seat at founding pricing. After this, Professional pricing applies to new sign-ups.

Reserve a founding seat

Why one project. Not a gallery of demos.

Demo projects prove you can follow a tutorial. Production projects prove you can ship, maintain and debug at 2am, and build systems that hold up under real users. Lucid is the latter.

  1. 01

    Patterns tested under production load

    Every RAG pattern, every safeguard, every escalation trigger in our voice agent builds ran first under real load in Lucid. Failure scenarios get found before they reach your callers.

  2. 02

    Responsibility, not delegation

    No framework generated Lucid's architecture. No team made decisions on a developer's behalf. Every design decision has a name and a reason behind it, we bring the same accountability to client builds.

  3. 03

    The same infrastructure you get

    Hetzner Falkenstein, Supabase pgvector, n8n orchestration, Lucid runs on the same stack we deploy for clients. We do not show demo infrastructure in the portfolio and hand you something different.

See it tailored to your clinic? Send a brief.

Tell me your current phone setup and call volume. I come back with a scope document, a timeline and an honest assessment of what is feasible for your configuration in 7 days.