Eight-plus years of spatial and vision systems, with LLM pipelines on top.
Photogrammetry, point clouds, imagery ML and WebGIS front-ends — plus modern retrieval and agent systems layered over them. Based in Solothurn, Switzerland; remote-first, onsite across DACH.
Computer Vision
Object detection and segmentation on aerial, satellite and industrial imagery. ADAS background at IAV — safety-critical perception, not demo notebooks.
Geospatial & Spatial Data
Photogrammetry, LiDAR and point clouds, PostGIS pipelines. Turning raw captures into data a product can actually query. Five-plus years at a German geodata company building spatial platforms and products that shipped.
WebGIS
OpenLayers and MapLibre front-ends — and the backends that feed them. Tiles, layers, cadastre data, the lot.
LLM & RAG
Retrieval over real documents, agents with tool use, model routing and evaluation. Grounded and cited, with the eval logs to prove it.
System Design
The architecture to run it all: APIs, data flow, cloud and Linux, MLOps, GPU when it earns its keep and not before.
Full-stack Delivery
FastAPI + React/Vite, Postgres, Docker. Shipped and running in production — not slideware, not a proof-of-concept that dies at handover.
Three ways in, depending on how defined the problem already is.
Scope, timeline and pricing are agreed per engagement — a short call is usually enough to tell which of these fits.
Discovery sprint
You know something is possible but not what it costs or whether it works. A short, fixed-length sprint that ends in an answer, not a maybe.
- •Feasibility on your real data
- •A working spike, not a slide deck
- •Architecture + honest cost estimate
- •Ends with a go / no-go
Consulting & advisory
You have the team, but not the depth in this particular corner — perception, retrieval, point clouds. I review what exists, cut the dead ends, and set the technical direction.
- •Architecture and code review
- •Build-vs-buy and feasibility calls
- •Technical direction your team executes
- •Single sessions or a short retainer
Project-based delivery
A deliverable you can name: a RAG pipeline, a detection model, a WebGIS tool, an automation. Zero-to-one, or hardening something that already limps.
- •Clear scope, timeline and price
- •You own the code and the infra
- •Handover docs, not tribal knowledge
- •Optional support retainer after
On Swiss compliance: engagements are billed through a licensed Swiss umbrella (PayrollPlus), so you get a clean supplier invoice with no risk of misclassification — the usual blocker for hiring an independent in CH is simply not there.
Spatial data & AI services
You capture the data. I make it deliver.
For survey, mapping, inspection and engineering teams sitting on imagery and point clouds: I turn raw captures into labeled training data, models trained on your classes, and classified results your clients open in a browser. Three steps — each one useful on its own, and you can stop after any of them.
Step 1 · Data
Domain-labeled training data
Annotation of aerial and street-level imagery and LiDAR point clouds — by an engineer who knows the difference between a wire and a branch.
- •Classes defined together, on your real data
- •AI-assisted pre-labeling on my own annotation platform — you pay for reviewed labels, not clicking hours
- •Delivered in your format: COCO, YOLO, classified LAS/LAZ
- •Per-dataset scope with a QC report
Step 2 · Model
A model trained on your classes
Detection, segmentation or point-cloud classification trained on that data and evaluated honestly — held-out metrics you can check, not a cherry-picked demo clip.
- •Imagery models and 3D point-cloud networks
- •Packaged as an API or container on your infra
- •GPU only where it earns its keep (Modal / RunPod)
- •You own the weights, the code and the data
Step 3 · Deliverable
Classified results as a service
You send captures, your client gets results: classified point clouds and extracted features, streamed straight to the browser — nothing to install on their side.
- •Corridor & asset classification: vegetation, wires, poles, road furniture
- •Browser-streamable output (COPC) — open a link, see the cloud
- •Per-project, or recurring as your captures come in
- •Runs on the same pipeline I operate myself
Why this works with one engineer: the annotation platform and the point-cloud streaming stack behind this track are systems I built and operate myself — that is what makes this viable without a labeling vendor, a separate ML team and a viewer license. Happy to demo both live on a call.
Proof
Systems I built, shipped, and still operate.
Not client logos — my own production systems, running on my own servers, where I am the one who gets paged. You can open every one of them right now.
Production RAG over a real corpus
/ask ↗Retrieval over posts and repos with embeddings, MMR reranking, grounded and cited answers, and per-query eval logging. Model routing through a self-hosted LLM gateway, so swapping providers is one environment variable.
GETCVAI.COM
getcvai.com ↗A dynamic CV engine — FastAPI, Postgres and an LLM agent API that tailors a CV to a specific role. Public, live, with a Chrome extension that auto-fills job applications from it.
PaperMind
live demo ↗A document vault with AI on top: OCR, extraction, vector search over pgvector, and a chat interface grounded in your own documents. Deployed, isolated per tenant.
Open-source LLM Gateway
github ↗An OpenAI-compatible gateway putting Anthropic, OpenAI and Gemini behind one endpoint — streaming, cross-provider fallback, per-key rate limits, budgets, and cost/latency/token tracking. It's what routes the LLM traffic in the systems above.
Demos
Watch them run.
Ninety-second screen recordings of the systems above — the real products in production, not mockups. Sound on for the narration.
PaperMind — documents in, answers out
Drop documents into a vault: OCR, auto-filing, extracted contacts and facts, and a chat that answers from your own files — with citations.
GETCVAI.COM — a CV tailored to the job
Paste a job posting, get a tailored CV and cover letter — an LLM agent pipeline over structured profile data, live at getcvai.com.
How it starts
Three steps, and you can stop after any of them.
A 20-minute call
You describe the problem. I tell you honestly whether it's a fit, and whether it's even worth building — sometimes the answer is no, and that's free.
A written scope
What gets built, by when, what it costs, and what "done" means. One page, no ambiguity, nothing hidden in a footnote.
We ship
Working software in your hands, in short increments you can see. Code, infra and docs are yours at the end — no lock-in to me.
Stack
The tools, for the people who want to see them.
Python · PyTorch · LLM & RAG orchestration · Anthropic / OpenAI / Gemini + open models · embeddings & vector search · eval pipelines
Computer Vision · point clouds & LiDAR · photogrammetry · PostGIS · OpenLayers / MapLibre
FastAPI · React / Vite · PostgreSQL · Redis · Docker · Linux · cloud & GPU (RunPod / Modal)
Get in touch
Where could this help you?
If you have a spatial, vision or document-heavy problem and nobody who owns the whole chain — that's the conversation. Twenty minutes, no pitch deck.
Capabilities overview — not a binding offer. Scope, timeline and pricing are agreed per engagement.