Project Case Study

Helios.
My personal AI
operating system.

What happens when a fintech product marketer with no engineering background decides that talking about AI isn't enough, and builds one instead.


Active Build · Phase 5
Helios · Current build status
Infrastructure
BuiltDocker platform + Portainer
BuiltSSD-backed persistent storage
BuiltTailscale VPN + SSH access
BuiltHomepage dashboard + Glances
BuiltHealth, backup + doctor CLI
AI Platform
BuiltOllama + Open WebUI
BuiltGemma 3 1B (general)
BuiltQwen2.5-Coder 1.5B (code)
BuiltQwen3.5 4B (reasoning)
BuiltModel registry + capability engine
ActiveMulti-model routing (Sprint 6.2)
Roadmap
NextAutomatic model discovery
NextProvider abstraction layer
NextPersistent memory
NextRAG + knowledge base
NextAgent runtime

The problem I couldn't ignore

I kept getting the same feedback. Different companies, different recruiters, different roles, but the same underlying message: strong on strategy, strong on fintech, strong on domain. But not technical enough. Couldn't build. Did not have engineering credibility. Could market AI but had not shipped anything.

That feedback is fair. It's also fixable. So I decided to fix it.

Not by taking a course. Not by getting a certification. Not by adding keywords to my resume. I bought a Raspberry Pi 5, put it on my desk, and committed to building something real until the feedback was no longer true.

"I didn't want to watch tutorials about AI. I wanted to understand what it actually takes to build an AI system that works in production."

The goal wasn't to become an engineer. I'm a product leader and I intend to stay one. The goal was to develop genuine fluency in AI systems, the kind that comes from building, breaking, recovering, and building again. The kind that changes how you think about product decisions, adoption challenges, and what it actually means to ship something that uses AI reliably.

What started as a learning project became something larger. It became Helios, a personal AI operating system that I'm building to grow alongside my career, and eventually to help run it.

Interactive routing demo
Ask Helios something.
Helios doesn't send every request to the same model. It classifies the job first, then routes it to the model that's best suited for the work. Try one of these examples or type your own.
01 · Input
Design a persistent memory architecture for Helios with RAG and provenance.
02 · Capability engine
Complex reasoning + architecture
03 · Model registry
Compare available local models
04 · Selected model
Qwen3.5 4B
Qwen3.5 4B
This request needs multi-step systems thinking and architecture tradeoffs, so Helios routes it to the strongest local reasoning model.
Front-end simulation of the routing logic used by Helios. The public portfolio page doesn't make live requests to my Raspberry Pi.

How it actually got built

I didn't come into this project with a plan. I came in with a Raspberry Pi, an SSH tutorial, and a list of things I wanted to understand: Linux, Docker, AI engineering, local LLMs, DevOps, software architecture. The phases emerged from the work itself.

Phase 1
Infrastructure
Raspberry Pi 5 setup, Raspberry Pi OS, SSH hardening, Tailscale VPN, VS Code Remote SSH, GitHub repository. Learning what it means to own a server rather than rent compute from the cloud.
Complete
Phase 2
AI Platform
Ollama, Open WebUI, Homepage dashboard, Glances monitoring. Docker and Docker Compose as the orchestration layer. External SSD migration for persistent storage. Three local models running fully offline: Gemma 3 1B for general chat, Qwen2.5-Coder 1.5B for code generation, and Qwen3.5 4B for complex reasoning and architecture decisions.
Complete
Phase 3
Recovery and Hardening
The SSD migration broke the Docker data root. Everything went down. I diagnosed the failure, rebuilt the storage architecture from scratch, and validated each service back to health.
Complete
Phase 4
BlueDemon CLI
Built a command-line interface with health, doctor, status, dashboard, and backup commands. Refactored from loose scripts into a proper modular Python package with clean separation of concerns. First time writing Python that felt like real software rather than glue.
Complete
Phase 5
Helios: The Intelligence Layer
The rebrand to Helios. Added the ask command. Built the routing engine, model registry, and capability engine. Helios now classifies queries by type and routes them to the right local model via the Ollama HTTP API. This is where the project became something more than infrastructure.
Current
Phase 03  ·  Recovery
The SSD migration
broke everything.
Diagnosing Docker data-root failures at midnight isn't in any tutorial. Rebuilding storage architecture from scratch and validating every service back to health is what operational discipline looks like — and it was the most useful thing that happened in this entire project.
System Architecture
How Helios works today
Current request flow from user input to model output
Userbd askHelios CLIhelios/cli.pyRouterhelios/routerModel Registry+ Capability EngineOllama API:11434/apiGemma 3 1Bgeneral · chatQwen2.5 1.5Bcode · generateQwen3.5 4Breason · architectREQUEST FLOWRaspberry Pi 5 · Docker100% LOCAL
Infrastructure
Docker Platform
Docker Compose · Portainer · External SSD · Tailscale VPN
AI Runtime
Ollama + Open WebUI
Gemma 3 1B · Qwen2.5-Coder 1.5B · Qwen3.5 4B · HTTP API · Local inference only
Intelligence
Helios Router
Model Registry · Capability Engine · Query Classification · Python package
Observability
Homepage + Glances
Service dashboard · Health monitoring · System metrics
Developer Interface
BlueDemon CLI
health · doctor · status · dashboard · backup · ask
Hardware
Raspberry Pi 5
8GB RAM · 500GB SSD · Raspberry Pi OS · VS Code Remote
Model evaluation
Evidence, not just architecture
Gemma 3 1B · Test: personal knowledge query
Failure observed
Asked what it knew about me. Returned confident personal information it had no basis for. Hallucination without any hedge, caveat, or signal of uncertainty.
Qwen3.5 4B · Test: open-ended reasoning
Failure observed
Entered an uncontrolled reasoning loop on an ambiguous architecture question. Kept generating without resolution. No natural stopping point.
System requirements produced by these failures
GroundingProvenanceExplicit unknownsReasoning budgetsRegression testsOutput constraints
Engineering lessons
What you learn when you build instead of watch
01
Docker is not just a convenience. It is a contract between the software and the environment. Moving data-root to an external SSD taught me what happens when that contract breaks, and how to rebuild it correctly.
02
Separation of concerns is not abstract. Naming a file router.py next to a package called router/ broke Python imports in ways that took real debugging to understand. Now I understand why the principle exists.
03
LLMs are not magic boxes. When I tested Gemma 3 and Qwen2.5 against identical prompts, I found hallucinated personal information in one model and uncontrolled reasoning loops in another. That experience changed how I think about AI product reliability.
04
API-first design matters. Helios communicates with Ollama through its HTTP API rather than shelling out to the executable. That decision makes the system testable, replaceable, and composable in ways that shelling out never could.
05
Recovery is the real curriculum. The SSD migration failure was more educational than any tutorial. Diagnosing what broke, rebuilding the storage architecture, and validating each service back to health is what production engineering actually looks like.
06
Routing is a product decision, not just a technical one. Deciding which model handles which query type, based on capability, cost, latency, and reliability, is the same problem product managers face when designing AI features for users.

On naming things

The infrastructure project has a second name: BlueDemonPi. Most people won't recognize the reference immediately, which is part of the point.

The origin of BlueDemonPi

Blue Demon, El Demonio Azul, was one of the most iconic figures in lucha libre, the Mexican wrestling tradition. He wore a silver and blue mask his entire career. Never took it off in public. The mask wasn't a costume, it was an identity. An alter ego built to do something extraordinary under a different name.

I named the project after him because I'm doing something similar. Danny Del Toro, fintech product marketer, is the public identity. BlueDemonPi is the alter ego, the side of the work that exists in the terminal, in Docker containers, in Python packages, in model registries. The side most people don't see.

Helios is the name I gave the intelligence running on top of that infrastructure. The sun. Something that generates light and makes other things visible. That felt right for what I'm trying to build.

The real lesson
I didn't learn AI engineering by reading about it.
I learned it by breaking things and fixing them.
The hallucinating model became a grounding policy. The broken Docker volume became a storage architecture. The import conflict became a module design principle.

Where Helios is going

The current version of Helios can route a query to the right local model and return a response. That is a foundation, not a destination. What I'm building toward is a personal AI operating system that can manage knowledge, automate workflows, and operate with increasing autonomy over time.

Each feature will be delivered as a complete vertical slice, code, tests, CLI integration, documentation, and a commit, rather than incremental fragments. That discipline is intentional. It is how production software gets built.

Local LLM infrastructure, Docker, Ollama, three models running fully offline
Model routing, query classification, model registry, capability engine
Automatic model discovery, HTTP-based Ollama API integration, dynamic registry updates
Provider abstraction, route to Ollama, OpenRouter, OpenAI, or Anthropic based on task and cost
Persistent memory, Helios remembers context across sessions
RAG and personal knowledge base, vector database, document ingestion, semantic search
Agent runtime, autonomous task execution, tool calling, workflow scheduling
Portfolio automation, Helios updates my website, drafts LinkedIn posts, generates GitHub changelogs
Career layer, resume updates, job application tracking, interview prep, long-term memory

"The end state is an AI that knows my work as well as I do and can help me operate at a level I could not reach alone."

What building Helios changed about how I think about AI

I'm not trying to become a software engineer. I'm a product marketing and growth leader with 14 years in fintech, and that is exactly what I intend to be. But the best product leaders in AI right now are the ones who understand the systems they are working with from the inside, not just conceptually, but practically.

Building Helios changed how I think about AI product decisions. When I discovered that one of my local models hallucinated personal information and another produced uncontrolled reasoning loops, I didn't just note it. I translated those failures into system requirements: grounding, provenance, explicit unknowns, reasoning limits. That's the same process product teams at Anthropic, Cursor, and every other serious AI company go through every day.

I know what it feels like to debug a Docker network issue at 11pm because a volume mount broke. I know why API-first design matters for a routing layer. I know the difference between a model that is fast and a model that is reliable. I know what it means to ship a complete feature slice with tests and documentation rather than a half-built prototype — that's a meaningful distinction.

The answer to 'can you build?' is now straightforward. But what changed more was how I think — about reliability, routing, failure modes, and what it actually takes to ship an AI feature that works in production rather than just in a demo.

Still building
Helios is not finished. It isn't supposed to be. It is a living system that will grow more capable as I build each feature slice, document each decision, and ship each commit. This page will too.
The GitHub repository is public. The architecture decisions are documented. The roadmap is honest about what is built and what is next. Come back as the system evolves, or reach out if you want to talk about what I'm building and why.
D
Danny Del Toro
Product Marketing & Growth Leader · San Francisco, CA
14 years in fintech and payments. Currently building Helios and owning product marketing for a $17.4B consumer card portfolio. Open to Director and VP roles at companies where AI is central to the product and the culture.