Bradley S. Isenbek
I'm your typical non-typical developer. While I've spent 15+ years architecting production systems that process billions of data points, what really drives me is the work I do in my garage lab, where I turn constraints into innovation and curiosity into working systems.
Philosophy
Case in point: I scored 60 Raspberry Pis from a failed business. Most people would resell them. I built a research tool instead: 40 of them working together, each one browsing the web on its own, privately and automatically. It's messy, it's scrappy, and it works.
I'm an integrator at heart. I dumpster dive for gear because working within constraints forces creative solutions. I run my own DNS server on a Frankenstein Linux box I built from salvaged parts. I host everything locally because I want to understand how it all fits together.
I bring the same mindset to production work. Find elegant solutions, question assumptions, and build tools that last. I've processed 4.9 billion data points, designed messaging platforms that stay up, and led classified government projects. Every one was a chance to build something that actually works.
I still host locally: this site is served from a box at home, not from a cloud. A Geiger counter and a Raspberry Pi turn radioactive decay into random numbers. Meatball is a home server built from other people's cast-offs that can see, hear and talk back. And there is a MOS 6502 simulated transistor by transistor from photographs of its die.
What I am building this year
Chips, from their photographs
A MOS 6502 chip, simulated transistor by transistor from photographs of its die. Then the two chips of an NES, done the same way. A console plays NES games in the browser, on faster models built from those simulations and checked against them.
The die photographs and netlists are other people's work: visual6502.org (CC BY-NC-SA) and Quietust's Visual 2A03 and Visual 2C02.
TerraPulse
A platform for measured climate and earth-science data. It holds readings from public instruments, never models or forecasts. AI assistants can ask it questions directly, through a public read-only connection. I am its architect. I also built tinychase, which runs small AI models on the same data in your own browser.
This site and its instruments
The pages here read real hardware on my network. There is a Geiger counter, an aircraft-tracking radio (ADS-B) and a GPS, Meatball's camera, and the logs of who tried to get in at the router and the web server. When a box is switched off, its page says so, and so does this card.
- Geiger counternot asked yet
- ADS-B and GPS buslive
- Meatball's cameraoffline, last heard 11 Sep 2026
- Door logslive
Free tools for the trades
Four field tools for trade workers. No account, no subscription, nothing uploads. They work for the person holding the phone and report to no one.
- Quote padline items in, a professional quote out
- Change-order padphoto, price, and a signature before the work
- Job photo stamperjob, time, and location burned into every photo
- Voice material padsay the list in the truck
Where the time went
Every role since 1997 on one axis, then six of them in a line each.
lib/resume.ts, from the resume dated 24 October 2026. Year resolution: each end of a bar is right to within a year.
Six of the roles, one line each
- 2024 to presentSysForge.aiFounder and AI Systems ArchitectMy own firm: AI consulting and development for enterprise clients. AI agents that do real work, and tools that bring AI into a team's daily work.
- 2022 to presentVictoryText, LLCArchitect and DeveloperHigh-volume messaging platform with FastAPI, DynamoDB, and multi-carrier integration.
- 2018 to 2022ConservativeConnector, LLCArchitect and DeveloperData systems processing 100M+ contacts and 4.9B data points with Snowflake.
- 2014 to 2018TransUnion, TruLookup DivisionSenior ArchitectSearch infrastructure over more than 10,000 data sources, a classified project with the United States Government, ML modeling.
- 2012 to 2014nextSource / PeopleTickerSenior ArchitectGlobal real-time labor-rate calculator with Java, Weka, and Apache Solr.
- 1997 to 2002PipeLive, LLCChief Technology Officer and co-founderReal-time collaboration software. 12 industry awards including PC Magazine's Winner's Circle.
Nine more roles between 2002 and 2026 are bars above and not told here, among them card payments, e-commerce, invoicing, and a case management system for a district attorney. The resume has every one.
What I work in
Mostly Python and TypeScript, PostgreSQL and Snowflake for data, Claude for the AI work, Raspberry Pi and ESP32 for the hardware, and Linux underneath.
The full list
- Languages
- Python, TypeScript, Rust, C and C++, Go, Bash, SQL, C# and .NET, Java
- Data
- PostgreSQL, Snowflake, DynamoDB, Redis, SQLite, Elasticsearch, Apache Solr, ETL pipelines
- AI and ML
- Claude and the Anthropic API, Ollama, Agentic frameworks, RAG pipelines, Vector databases, Hugging Face, PyTorch, LangChain, Entity resolution
- Infrastructure
- Linux, Nginx, Docker, AWS, Kubernetes, FastAPI, High-availability design
- Hardware
- Raspberry Pi, ESP32, Arduino, LoRa, Edge TPU, Custom protocols
- Tools
- Git, Vim, Playwright, Jupyter, Prometheus, Grafana, CI/CD
Writing
Three field notes from building Meatball, each written up as it happened, with the numbers from the run that produced them. They are on the Meatball page.
How this site is built
Two of us build this site, and the split is plain.
Bradley Isenbek
Grand Rapids, Michigan
- Decides what gets built
- Checks the result on the live site
- Puts his name on what goes out
Claude
Claude Code: checking
- Writes most of the code, in sessions logged at /ai-pilot
- Writes and maintains the data pipelines, which run on a timer
- Runs the deploys
Claude is Anthropic's model, working here as Claude Code: a collaborative development partner, not just a code completion tool. The sessions are counted in public. The AI pilot record shows at least 5,152 of them on 10 models, from 8 March 2026 through 30 October 2026, across every project on this machine and not this site alone. Recording began on 7 July 2026, so the months before it are incomplete. The version in the footer counts runs of the deploy script: this is number 508.
Two rules keep it honest. A number a pipeline counted does not appear without the file it came from and the period it covers. And work is reviewed adversarially before it ships: a second pass reads it with one job, to find what is wrong.
Signed 2 October 2026Bradley IsenbekClaude
Sessions and models in the text: ai-pilot-data.json, generated 30 October 2026. The run number is this build's version.
Open to full-time roles in AI systems and data architecture.