Lukas Oberbichler

Lukas
Oberbichler

Software shaped around your work.

I’m a software engineer building practical AI tools and workflow software, from the first conversation through delivery.

Let’s talkExplore my work

Remote, with occasional agreed on-site collaboration if requested

Lukas Oberbichler with a city skyline in the background

Letter

From a business need to something you can try.

I’m building Letter to help business owners and employees explore software for the work they actually do. Describe a workflow, try an example, and discuss what should change.

Try Letter
Store sales & invoices
Store sales & invoices

Real screenshots from Letter’s public demo. The example business records are fictional.

Explore the architecture

Built for agent-led improvement

I’m following Langfuse’s and other's concept of the AI Engineering Loop. This way of building applications enables agents to inspect, evaluate, experiment with, and improve. Human feedback helps define what better means.

  1. A scoped request

    The problem, intended outcome and acceptance criteria are recorded together.

  2. An isolated build

    An agent works on a candidate without replacing the usable version.

  3. Independent review

    A separate reviewer examines the exact candidate and reports findings.

  4. An evidenced release

    Qualification connects source, checks and the observed deployed result.

Pieces I’ve implemented

Observable behavior
Trace export, scores and authenticated readback provide evidence that can be inspected by people and agents.
Explicit change boundaries
Revision-bound briefs, isolated candidates and bounded review-and-repair cycles keep a requested change inspectable.
Checks tied to the result
Release qualification connects the candidate’s source to browser checks and matching observations. Incompatible changes can be held for review.

The AI Engineering Loop

Live activity reveals what needs work. Development turns those observations into tested changes, then deployment starts the next cycle.

  1. TraceOnline

    Capture what the application does in real use.

  2. MonitorOnline

    Find patterns and problems; add human feedback.

  3. Build datasetsOffline

    Turn useful examples into repeatable test cases.

  4. ExperimentOffline

    Try changes to prompts, models and code.

  5. EvaluateOffline

    Compare results against defined quality criteria.

Deploy
Adapted from Langfuse’s AI Engineering Loop. View the original diagram

These are implemented building blocks. The broader dataset, experimentation and evaluation workflow is still being developed.

The Langfuse approach

Engineering, with people

I like taking a problem from a conversation to something people can actually use. My work has taken me from embedded automotive software to museum installations, construction tools and AI applications. Working directly with customers — including years of IT support for seniors — taught me to listen carefully and explain technical choices clearly.

Earlier work

Bosch
Embedded ECU software in C and Matlab/Simulink, tested in prototype vehicles and hardware-in-the-loop simulators.
Inomedia
Ruby on Rails museum software and Linux video kiosks, from inherited code to on-site installation and handover.
Build Informed
C# and F# Revit extensions for construction planning, legacy-code migration and workflow automation.
Tribe
Co-founded a creator-funding product and helped build a music-licensing prototype presented at Red Bull’s London offices.
Read the organizer’s account

What we can work on

  • Automate a repetitive business workflow
  • Build an AI-powered tool or product feature
  • Improve an existing application or integration

Work together

Looking for someone to strengthen your team, or take a project from problem to delivery? I’m open to both conversations.

Discuss a roleDiscuss a project
Book 15 minutes

Remote, with occasional agreed on-site collaboration if requested