Skip to content
Alain Lorenz

Alain Lorenz — CTO & co-founder · AI & MLOps · Geneva

Full stack, quite literally:from the phone down to the GPU.

I build complete products and run their AI models in production — from raw data to GPU inference. Co-founder and CTO of JobyPepper and Ozly.

See the projectsGet in touch
Trace of one search, from screen to GPU1 request · 6 layers · 1 developer
  1. clientapp.tap("Search")
  2. apiPOST /search
  3. queuerabbitmq.publish(search)
  4. searchvector.search(k=20) + filters
  5. modelcamembert.classify → llm.answer
  6. gputriton.infer(gpu:0)
#about

The CTO who still commits.

As co-founder and CTO, I have led teams of up to ten people without ever letting go of the keyboard: I am still the top contributor to the API I started in 2016. Since 2023, my main playground has been AI in production — training, deploying, serving and optimising models that answer real users. My favourite subjects: LLMs, RAG and fine-tuning. All from Geneva.

Yes, a CTO who commits. The species is not extinct.

#path

The path of a request

I built every layer it crosses. Half of the trip happens on the AI side.

  1. The screen

    app.tap("Search")

    React Native / Expo mobile apps, including a complete business app, taken all the way to the stores: iOS and Android builds, submission, listing copy. React web interfaces for admins, clients and freelancers.

    It all starts in a pocket.

  2. The API

    POST /search

    Node.js / Express, business workflows, and a financial dispatch engine: a single client payment automatically becomes every freelancer’s invoice, the fund transfers and the payouts.

    Where the money flows. No pressure.

  3. The queue

    rabbitmq.publish(search)

    RabbitMQ between the API, workers, scheduler and real-time services. Each service does one thing, asynchronously.

    Everything gets through eventually. In order.

  4. The model

    camembert.classify → llm.answerAI

    Fine-tuned CamemBERT classifiers, sourcing and matching agents, MCP servers that plug LLMs into real business data.

    The part everyone talks about.

  5. The GPU

    triton.infer(gpu:0)AI

    Inference served by Triton on GPU, CUDA containers, GPU cost divided by four.

    The part that gets expensive when nobody is watching.

#lifecycle

A model in a notebook is a promise. In production, it is a service.

MLOps is everything in between. I cover the whole cycle.

  1. 01

    Collect

    Scheduled pipelines that aggregate, clean and deduplicate data from multiple sources.

  2. 02

    Train

    Dataset building, CamemBERT fine-tuning to classify by sector and contract type.

  3. 03

    Package

    Versioned models, CUDA Docker images, published to a container registry.

  4. 04

    Serve

    GPU inference through Triton, an embedding service on Kubernetes, vector search.

  5. 05

    Operate

    CI/CD, observability on every service, continuous optimisation: GPU ÷ 4, embeddings cut down to 512 dimensions, a memory leak hunted down and fixed.

Step 5 never ends. That is expected.

#work

How I work

Projects run end to end with AI, and a human at every checkpoint.

Pipeline of a project, from idea to productionin production
  1. notionThe project is declared in Notion.human
  2. reviewIt is reviewed and approved by a human.human
  3. buildLLMs build it, plugged into the code and its context.AI
  4. verifyI verify, test and fix.human
  5. deployShipped to production.CI/CD
  6. monitorErrors flow into Sentry, readable by the agents.agents + human

My real speciality is what happens after the release: monitoring, diagnosing, stabilising — in record time.

The agents read Sentry before my first coffee. The fix often ships before it gets cold.

Start from the need

  • Before the first line of code: understand the business, who will use the tool, to do what, under which constraints.
  • Derive the user journeys from that, then the UX/UI. The interface follows the need, not the other way round.
  • An in-house design system so the whole product speaks one language.

Leading

  • Technical leadership of a team of up to ten people.
  • Architecture decisions, delivery planning, product / tech trade-offs.
  • Preparing innovation funding applications.
  • Turning an operational need into a tool, and an AI experiment into a service that runs.

Working with AI, not next to it

  • Coding agents in my daily workflow, with versioned instructions in every repository.
  • MCP servers built on three products to plug LLMs into business data.
  • Production errors are readable by the agents: diagnosis and a proposed fix, right away.

This site was designed with two AIs. I wrote the brief, settled their disagreements and proofread every sentence. That is called management.

#projects

Projects

Technical co-founder · CTO · since 2016

JobyPepper

The platform that connects companies with freelancers for short missions.

  • Web, mobile, API and infrastructure: the whole product ecosystem.
  • Two mobile apps published on the iOS and Android stores, from build to listing: versions, submission, compliance fixes, store copy.
  • Payment dispatch, not just a “pay” button: the client pays once, the platform generates every freelancer’s invoice, splits the funds and triggers each payout. Automatically.
  • AI agent for candidate sourcing and matching, connected through MCP.

Technical co-founder · CTO · 2023 · on hold

Ozly

AI applied to job search.

  • Full pipeline: multi-source ingestion, CamemBERT classification, vector index.
  • Semantic search and conversational RAG, in an embeddable widget.
  • Optimised GPU inference: cost divided by four.

On the side

Personal project · 2026

Origin AI

Videos of real workplaces turned into training data for robotics.

  • Chunked video upload, ffmpeg processing, LLM tagging.
  • Automatic face blurring on GPU, dataset creation pipeline.

Side project · 2026

EV Comparator

After how many kilometres does electric become cheaper?

  • Daily open data ingestion, cost-per-kilometre model.
  • Resale value shown as an estimate, never as the truth.
#play

Off-piste.

I am not just a geek. Above all, I am someone who has fun — on the water, on the snow, on a court, at the stove, behind a lens or in front of a screen.

  • Board sports

    • Skiing
    • Wakeboarding
    • Surfing
    • Foiling

    Same rule as production: as long as it keeps moving, it stays up.

  • Racket sports

    • Anything played with a racket

    One request, one response. Slightly more out of breath.

  • Creative

    • Cooking
    • Photography

    A recipe is a pipeline whose output you eat. A photo is a render you cannot re-run.

  • Pop culture

    • Video games
    • Cinema
    • And everything else

    Yes, I watch the cutscenes.

#timeline

Timeline

  1. 2016

    JobyPepper — technical co-founder, CTO

  2. 2023

    Ozly — technical co-founder, CTO · on hold

  3. 2026

    Origin AI — personal project

Languages — French · German · English

#contact

A project, a role, a question about RabbitMQ?

The fastest way is Instagram.

I hate LinkedIn.

Based in Geneva.