Justin Boehnen

Justin Boehnen

Senior Software Engineer | Platform Architect | AI Engineer

What I Do

I architect data pipelines that wrangle massive, messy data from the real world. I integrate AI and ML into workflows to improve efficiency and accuracy, from LLM-powered analysis to custom models that save hours of manual work. And I ship the full stack: backend APIs, web apps, mobile.

Currently

I'm the Cloud Development Lead creating land management software for the largest private landowner in the US. I own the platform architecture, backend APIs, infrastructure, and production operations for a microservices system serving mobile, desktop, and web clients. I'm currently integrating AI, MCP, ML, and LLMs to improve user experience and workflow efficiencies across the board.

C# .NET TypeScript Python Azure PostgreSQL Docker LLM Integration MCP

I build world-class APIs and AI integrations

BrainGap Live

Demo

Adaptive Assessment Platform

The Problem

Building quality assessments traditionally requires psychometricians and months of calibration. Most AI quiz tools just generate questions with no validation.

My Solution

A public API that delivers rigorous adaptive assessments to any application. Under the hood: Claude generates structured content, OpenAI embeddings deduplicate concepts, and every response calibrates item parameters using the same 2PL IRT model behind the GRE and GMAT.

How I Use AI

I use AI to be the expert where I'm not. But I don't blindly trust it. Every AI-generated item flows through grounded feedback loops, validated by real data, not model confidence. Poor performers self-identify and flag for review.

.NET Aspire Blazor PostgreSQL pgvector Claude API OpenAI Embeddings OAuth

I make complex data accessible

ClipMap Live

Demo Source

Zero-Backend Geospatial Exports

The Problem

The maker and art map market needs SVG exports of geographic regions for laser cutting, CNC, and wall art. The pipeline for getting these files is not well defined, laborious, complex, and slow.

My Solution

A website where you select any location, pick your layers, and download production-ready SVGs in one click. Behind the scenes: Node.js and mapshaper pre-processed 95GB of OpenStreetMap into 157k GeoJSON tiles across 5 LOD tiers. All clipping, projection, and SVG rendering runs client-side.

Why No Backend

Pros: Zero compute costs, infinite scale, no cold starts, globally distributed via CDN. Tradeoffs: Data must be pre-processed, can't support arbitrary queries, tile updates require full regeneration.

What's Next

Roads and boundary layers in progress. Exploring vector tile formats for smaller payloads.

TypeScript Next.js React Leaflet Node.js mapshaper GeoJSON Cloudflare R2

I train models for the real world

SatSuite

Satellite Imagery ML Pipeline

The Problem

Satellite imagery analysis requires specialized ML pipelines. Off-the-shelf models don't generalize well to aerial perspectives and specific land cover categories.

My Solution

Drop in coordinates and get instant land cover analysis. Built by fine-tuning SegFormer-b0 on OpenEarthMap for 8-class segmentation (mIoU: 0.596). Training in ~45 min on a free Colab T4. Deployed to HuggingFace Spaces for free inference.

Challenges

Class imbalance in satellite data, inconsistent imagery quality across regions, and managing inference costs for large-scale analysis.

What's Next

Change detection and object counting models. Train on better hardware, more data, longer. The goal is production-grade models for real-world geospatial analysis.

Models

SatLens Live Land cover segmentation Demo Source
Python PyTorch Transformers SegFormer Gradio HuggingFace OpenEarthMap