Selected Work
The work.
A selection of things I've built, why I built them, and what they were actually trying to solve.
AI-Augmented SDLC
The most important shift in software development right now isn't the tools. It's the process.
When code becomes a commodity, the leverage moves upstream: to how you frame the problem, how you interrogate dependencies before a single line gets written, and how you maintain the taste and judgment to know when the output is actually good.
I've built a complete AI-augmented software development lifecycle around this insight. The methodology is built on four principles that sound simple and are harder than they look in practice:
Decompose the problem.
Break it into its real constituent parts before touching the tools. Not the surface problem, the real one.
Describe the desired outcome.
With enough precision that anyone working toward it knows when they've arrived. Not a direction. A destination.
Apply taste and judgment.
Evaluate what comes back honestly. The model will give you something. The question is whether it's the right something.
Hold the line.
Until the work is actually right, not just done. This is the step that most AI-augmented workflows skip, and it's where the quality gap opens up.
At the center of this methodology is Codex Designer, a cross-platform CLI tooling workflow that acts as the bridge between taste and judgment on one side and executable code on the other. It is the operational manifestation of the principle that what you know before you ask determines almost everything about what you get back.
The result is dramatically fewer iterations, dramatically higher first-pass quality, and output that holds up in production rather than just looking good in a demo.
This methodology is what New Era Systems runs on. It is available as a consulting engagement, both as a direct development service and as a process that can be learned and embedded in your own team.
Automated Assessment Tooling
A fully automated prospect and client intelligence pipeline.
A respondent moves through a structured, scored assessment. The system runs parallel web research in the background. A branded narrative report is generated via AI and delivered to the recipient's inbox in under five minutes. Zero human intervention between submission and delivery.
The automation is the easy part to describe. The harder and more interesting part was designing the assessment rubric so that the output was genuinely useful rather than generically reassuring. Scoring without judgment is just math. The goal was insight: a report that told someone something true and specific about their situation, not a document that made them feel good about having filled out a form.
The architecture is applicable across consulting intake, risk evaluation, AI readiness assessment, and any context where the need is to turn structured input into a meaningful, personalized output at scale.
DialogDrive
A state-aware sales outreach automation system with enough intelligence to understand where a conversation is in its lifecycle and respond accordingly.
Most outreach automation sequences emails. DialogDrive reasons about the state of a relationship: where the prospect is, what they've engaged with, what the next meaningful touchpoint actually looks like. The difference is the difference between a system that follows a script and one that reads the room.
Funded by a strategic partner in exchange for lifetime use and vertical exclusivity. Currently in active deployment.
Multi-Agent AI Debate System
Open SourceA system where multiple AI models debate a problem. Each model takes a position, responds to the others, and is evaluated by a referee model for logical convergence. The goal is to surface where models agree, where they diverge, and why.
The interesting insight the system produces isn't the winning argument. It's the map of where the disagreement actually lives. That turns out to be useful information.
The initial version was built in approximately 30 minutes using Codex, followed by a couple of days of refinement to turn the core idea into something publishable. Published because the architecture deserves to be in the world.
OpenResearch Harness
Open SourceAn open research harness for running structured, repeatable research workflows with open source software and open weights models.
One-shot prompting is fine for a draft. It is not a serious research process. OpenResearch was built to make retrieval, synthesis, orchestration, and iteration reusable instead of improvised every time a new question shows up.
The value is in turning research from an ad hoc conversation into durable infrastructure: workflows that can be rerun, inspected, improved, and applied across new domains without starting from scratch.
Study with Heffy
A full AI coaching platform built for HESI Med-Surg 2 exam prep. This started as something I built for my daughter. It became a real product.
The platform includes a personalized AI coach with her own avatar and persona, adaptive targeting that tracks weak areas and focuses sessions where they're actually needed, RN-reviewed practice content, dynamic on-the-fly visuals for tricky concepts, and downloadable study guides generated from every chat session.
The coaching philosophy is in the product: the point isn't to pass the test. It's to build a nurse. Answers that stick on exam day and on the floor. That distinction shapes everything about how the coach responds.
Subscription at $10/month. Auto-stop after three months of inactivity, because charging people who aren't using something is its own kind of wrong.
Ordo.Chat
A fully released platform for building and deploying production AI agents.
Ordo was built around a simple idea: agents become useful when they can operate inside the real shape of work. That means grounded context, tool access, visual explanation, and a clean path for human takeover the moment the situation calls for judgment.
The platform supports configurable agents, retrieval over uploaded knowledge, connected tools and APIs, chart and diagram rendering, structured insight capture, and deployment across web and collaboration surfaces. The point was never to ship another toy chat interface. The point was to make agent systems usable in production.
Beyond the Work
Not everything worth building is a client engagement.
I've hosted karaoke for years. There is real craft in reading a room, building energy across a multi-hour show, and knowing when to get out of the way and let the moment happen. I'm training an apprentice to eventually take over hosting duties — which is, if you think about it, a reasonable summary of how I approach most things I care about. The goal is always to make the people around me more capable, not more dependent.
When Karafun 3 launched without porting over years of song history from Karafun 2, I did what any reasonable person would do: built a full application out of spite. automaticmayhem.com/sing interfaces with Karafun, adds try lists, detailed singer history, AI-powered song suggestions, and real-time queue statistics so every singer always knows when they're up next.
studywithheffy.com is the one personal project on this site, and it belongs here because it is one of the clearest expressions of what I actually believe. It started as an AI coaching platform I built for my daughter, who was preparing for her HESI Med-Surg 2 nursing exam. It became a real product: a personalized AI coach with her own avatar, RN-reviewed content, adaptive weak-area targeting, dynamic visuals for tricky concepts, and downloadable study guides from every session. The coaching philosophy is in the architecture: the point isn't to pass the test, it's to build a nurse. $10/month, auto-stop after three months of inactivity, because charging people who aren't using something is wrong.
Outside of client work, I volunteer with small business coaching programs, show up at pitch competitions, and work with organizations focused on young entrepreneurs. The principle is the same everywhere: the answer someone else gives you is useful once. The understanding you build yourself is useful every time after that.