Platforms and backends
APIs that hold when the timetable changes at four in the morning.
Architecture and implementation for API middleware, microservices and internal developer platforms, using Node.js, TypeScript and Java.
Software engineering studio · Vienna
Systems that carry the load.
High-performance Node.js and Java backends, developer platforms and edge AI for services that need to stay up.
Network study
Shipped for
Software engineering. Three areas of focus.
Edge AI and computer vision
Inference where the data already is.
I wrote the Node.js bindings for Hailo accelerators and Raspberry Pi camera modules, so the model runs next to the sensor and only the answer travels.
Cloud and reliability
The work that keeps the pager quiet.
Kubernetes, CI/CD and observability for services in production, with profiling and migrations guided by the system’s constraints.
Selected work
Where downtime makes the news.
01 ÖBB
Ticketing for a national railway
Solution architect and technical team lead. Together with the team, we developed API middleware and high-performance applications in Node.js and Java.
tickets.oebb.at
Westbahn corridorEngineering notesfor ÖBB Architecture / technical leadership
Why this approach
Separate orchestration from service logic.
A layered middleware boundary gives API composition and individual microservices distinct responsibilities. Across Node.js and Java, that separation makes the request path easier to reason about, profile and evolve.
- Context
- Booking and payment for tickets.oebb.at, the online ticketing service of a national railway.
- My role
- Solution architect and technical team lead, connecting architecture decisions with the team’s implementation work.
- Engineering
- Performance-focused Node.js and Java applications, API middleware and clearly separated microservice layers.
02 Mercedes-Benz.io
One front door for hundreds of services
A Backstage developer portal that gives teams a place to find services.
mercedes-benz.io
Service dependency graphEngineering notesfor Mercedes-Benz.io Finding a service
Why this approach
Give service discovery a shared front door.
A Backstage portal makes finding a service part of the developer workflow. Teams can start with a shared catalog, then use a conversation for the questions the catalog cannot answer.
- Context
- Hundreds of services to navigate. Finding the right one can start with a question in Slack.
- Approach
- A developer portal built on Backstage provides a front door to the service landscape.
- In practice
- Teams have a place to look for a service before starting a conversation to find it.
03 Johnson & Johnson
Shared data, findable across a giant
Shared data is only useful when the right dataset can be found across departments.
Federated indexEngineering notesfor Johnson & Johnson Across departmental boundaries
Why this approach
Treat discovery as part of sharing.
Storing a dataset does not make it useful to another department. A discovery layer has to cross the same boundaries as the people trying to find it; that is the purpose of the federated index.
- Context
- The right dataset can sit several departments away from the person who needs it.
- Approach
- Make shared data findable across those departmental boundaries.
- In practice
- Reach matters alongside storage: a dataset has to be findable from another department.
04 dofbot In house
A robot that grasps what it cannot see
Detection, depth and grasp run on the robot, with no cloud in the loop. The camera cannot see into the gripper’s working pocket. The final approach has to bridge that gap.
From sight to grasp.
SchematicFinish the reach from what was seen.
The jaws work between 14 and 24 cm. The arm uses the previously observed target to finish the approach.
Inside the decisionfor dofbot Reach / sight / final approach
Why this approach
Stop. Observe. Correct. Then commit.
Navigation gets the robot close, but its tolerances are wider than the grasp pocket. Inference also takes time. The robot stops for a fresh observation, makes a small correction, and verifies the target before a short, guarded straight approach.
- Constraint
- The safe grasp pocket is 14–24 cm from the base. The eye-in-hand camera cannot see the target in that pocket.
- Handoff
- In the taught approach, verify the target while it is visible, then use odometry for the final calibrated straight segment.
- Guardrails
- No steering after the last verified camera observation. The collision guard must be enabled, and the blind advance is capped at 22 cm.
What that adds up to
From platforms
to play.
Production systems, public code and ideas we build ourselves.
Read the code before you hire us.
Inside the open source
A frame’s journey.
Camera → on-device inference → a useful decision.
01 / Camera
One frame. Two useful views.
Keep a JPEG stream for people to see and a smaller pixel stream for the model to process. Each gets the detail it needs.
Light Image data
Explore picamera.jsimport { builder } from '@nodify_at/picamera.js';
const camera = builder()
.jpeg(1280, 720)
.rgb(640, 360)
.fps(30)
.build();02 / Inference
Let the device do the seeing.
Prepare the image for the loaded model, then run it on the Hailo accelerator. The result comes back to the application as model output.
Prepared image Model output
Explore hailo.js// Model loaded; image prepared.
// inputName comes from model metadata.
const outputs = await device.infer({
[inputName]: resized
});03 / Decision
Seeing something is only the start.
The application turns detections into a useful response. Freshness, confidence and the rules of the task decide what happens next.
Detections Application response
Read the tracking examplewhen a detection arrives
if the frame is stale
wait for a fresh observation
if the target is uncertain
observe again
otherwise
pass the target to the applicationAn example of the rules around a model.
Illustrative flow · Explore the steps to see where each library fits.
nodify-at/svelte-formify
Form handling for Svelte with full type inference.
Apache 2.0 nodify-at/gpt-pull-botThe first review pass a human should not have to do.
MIT nodify-at/hailo.jsHailo AI accelerators, without leaving JavaScript.
C++MIT nodify-at/picamera.jsPi camera modules with hardware acceleration intact.
C++MIT nodify-at/redux-mappersState and dispatch to props, without the boilerplate.
MIT All repositoriesIncluding the experiments that never became libraries.
Small on purpose.
You work with Hasan Özdemir in Vienna, from software architecture to implementation. Nothing sits between the conversation and the commit.
Studio notes 01
dofbot · Perception & control
Carry the measurement through.
The camera loses sight of the target before the gripper can reach it. In dofbot’s taught approach, the target is verified while it is still visible, then the base follows a short, guarded straight path.
The final grasp keeps the small position difference measured at that last check. It adjusts the grasp target; the base’s calibrated path stays the same.
The dofbot projectA handoff should preserve the information the next step still needs.
Tell us what it has to carry.
Also from nodify
Dimwood
A cosy iOS game about leading small creatures home through a dark wood. Built here start to finish, which is the only brief where we were also the client.
- Swift
- SwiftUI
- Metal
- App Store
