Draft portfolio page

Building practical AI systems

Hands-on experience turning AI into agents, automations, local model experiments, browser workflows, knowledge bases, and reviewable tools tied to real software projects.

Review examples See professional direction How the workflow is structured See local model benchmarking

From curiosity to infrastructure

The thread through the work is simple: treat AI as software infrastructure, not as a novelty layer.

Start

An open-source agent framework and the first agent environment

My AI journey started when I heard about an open-source agent framework and wanted to understand it by building with it directly. Around mid-March 2026, I spun up a DigitalOcean VPS, secured access with Tailscale, connected Anthropic models, and created my first agent environment.

That first setup quickly moved past experimenting with prompts. It became a practical learning environment for agent workflows, scheduled jobs, Telegram-style interactions, memory, logging, auth flows, and small web applications.

Shift

Structured AI workflows in Hermes

I later moved into Hermes as a more structured agent environment. That gave me a place to work with profiles, toolsets, model routing, Obsidian-backed project memory, browser automation, and clearer review gates.

I have also experimented with local model hosting on an Apple Silicon Mac Mini, comparing hosted models with local serving paths and learning where each one fits.

Projects that show the shape of the work

These examples are intentionally high level. The goal is to show practical systems thinking without exposing private data, credentials, family details, or live endpoints.

Applied workflow

Manitoba Cottage Search

A FastAPI and HTMX listing-review app with structured intake, browser-backed extraction, scoring, and provenance-aware persistence. It separates import, review, and status so new listing candidates can move into a real application instead of disappearing into spreadsheet workflows.

  • Structured intake API with deduplication and audit trail
  • Browser-backed saved-search extraction with a freshness gate
  • Review tracking design that keeps rejected listings explainable
Read the case study
Agent architecture

Hermes workflows and model routing

A structured AI working environment using profiles, toolsets, model selection, documentation, and review gates. This is where I started treating AI work as an operating model: different lanes for planning, coding, evidence gathering, local experiments, and public-facing writing.

  • Profile-based role separation
  • Obsidian-backed durable context
  • Hosted and local model tradeoff testing
Read the Agent Workflow page
Local AI lab

Local model benchmarking

A practical benchmark pass on Atlas, a Mac Mini M4, comparing Ollama, MLX research paths, and oMLX serving. The work focused on generated-code correctness, memory ceilings, model unloading, and where local models fit inside real workflows.

  • Apple Silicon local serving with Ollama, MLX, and oMLX
  • Python code-generation benchmark checked with pytest
  • 17.8GB Metal memory cap documented as a real constraint
Read the benchmark page
Agent-built web app

Student Assignment Tracker

A mobile-first education tracker built with FastAPI, SQLite, and HTMX through an agent-assisted development workflow. It models classes, schedules, assignments, semesters, and grades, then documents the planned agent-facing API separately from what has shipped.

  • Concrete product model with real workflow logic
  • Agent-assisted build process using FastAPI, SQLite, and HTMX
  • Designed and planned agent API layer, not implemented
Read the case study

Practical AI is mostly systems work

01

A prompt is not a system.

Useful AI needs state, logs, scripts, review gates, and clear handoffs.

02

The boring plumbing matters.

OAuth, timezones, delivery channels, file ownership, and failure states are where demos become real software.

03

Model choice depends on the job.

Cheap models, strong models, local models, and deterministic scripts each have a lane.

04

Human review is a feature.

Especially for public communication, workflow changes, private data, and anything with external side effects.

Applying it to enterprise-relevant AI work

I am now translating this hands-on experience toward professional software and enterprise AI use cases: internal document assistants, support-ticket triage, workflow automation, AI-assisted developer productivity, and responsible Microsoft/Azure/Copilot adoption.

My strongest angle is practical implementation. I understand web development, real workflows, messy integration points, and the need for governance. The goal is not to add AI for novelty. The goal is to improve useful work safely.