What South Denver’s AI Builders Are Actually Shipping

AI demos are easy to find. Working systems—with constraints, customers, infrastructure, and a point of view—are much more interesting.

At the August South Denver Let’s Vibe meetup, builders brought projects that went well beyond chatbot wrappers: an agentic software-development harness, creator tools designed to preserve human authenticity, experimental interfaces for education and media, and a CRM built for AI agents. The projects were different, but the same question kept surfacing: now that AI can do meaningful work, how do we make that work reliable, useful, and worth trusting?

The shift from chat to infrastructure

Shea Hawkins demoed OpenEngine, an open-source harness designed to keep coding agents working across delegated software projects. The goal is not to ask an assistant for one patch at a time. It is to build a repeatable system for provisioning work, implementing changes, running deterministic checks, collecting specialized reviews, and routing the results back through another iteration.

OpenEngine currently runs locally with SQLite and is moving toward a more durable setup built around Temporal, Postgres, and an always-on Mac Mini. Its review model was especially interesting: multiple reviewers inspect security, performance, and other concerns, then a re-ranker tries to surface useful signal instead of handing a person four repetitive walls of feedback.

The intended daily experience is simple: wake up, review a small batch of merge requests, make the decisions that require human judgment, and let the system refill its available capacity. That is a much more mature framing of agentic development than “AI writes code.” The hard problem is orchestration—budgeting, pacing, verification, and deciding when a person must step in.

Using AI to remove production friction—not the person

Greg Narain showed two creator products built around a different idea: AI should make it easier for a real person to create, not replace the person with synthetic output.

Zealous ingests recordings, podcasts, and live conversations, then produces editable transcripts, engagement heat maps, highlights, and audiograms. Its rendering workflow uses Remotion to preview video in the browser and distribute frame rendering across cloud functions.

Greg also demonstrated Reactionary, a prototype built in about a day. A creator plays a clip, records a reaction, and the system automatically preserves the few seconds of context needed to understand the response. The user can edit the resulting clips without opening a traditional video editor. An audiogram mode strips away the source video and renders the response as transcript-driven media.

His pitch captured the product philosophy: “There’s no AI in this. It’s actually you.” That matters as platforms increasingly penalize low-effort, AI-only content. The durable opportunity may not be generating more content. It may be removing enough friction that knowledgeable people can publish their own ideas consistently.

Experiments are still one of the best ways to learn

Ken Macy brought a rapid-fire collection of experiments that showed how broad the current toolset has become.

Ken’s experiments included encoding images and text into variations of the Amen Break, then decoding the audio back into the original content. He also showed Claude controlling Blender through MCP to create a character, animate a walk cycle, and export assets for a Three.js portfolio. A separate NotebookLM experiment transformed an essay about The Mighty Ducks into a surprisingly serious podcast and leadership framework.

Ken's most personal project was a local-first Japanese-language phone companion designed for a family member. Private context stays on an Nvidia Jetson, while voice and print outputs create a familiar daily interaction. It was a useful reminder that “local AI” is not only an architecture preference; sometimes it is the difference between an idea being appropriate and being invasive.

Eric Schneider shared an experiment that turns familiar movie stories into AI-generated commentary. He gives the film script to his AI system to write a summary, imports that to NotebookLM which will produce a podcast and supporting visuals that reinterpret the story through unexpected lenses—such as leadership, trauma, competition, and personal growth. His demo used popular movies, and the interpretations had us all laughing and wondering why so many of the interpretations were trauma focused!?

Matt Bernier demonstrated a prototype that turned a dry-erase board into a guided learning interface. A student types a question, the system translates it into a structured Wolfram Alpha query, and the response appears as handwriting on a digital whiteboard. Answers can lead into follow-up questions, encouraging exploration instead of one-and-done copying. He also showed off an update of his calendar tool, a10d.info. In a live Demo, his AI assistant created a contact from a new attendee’s name and email, and stored notes from the Granola recording of the meetup.

The tools worth watching

Several supporting tools and shifts came up throughout the night:

ACP is emerging as a way for development harnesses to work across Claude, Codex, Gemini, and other command-line agents.
Remotion makes programmatic video production approachable to React developers and can scale rendering frame by frame.
NotebookLM continues to produce surprisingly deep podcast-style audio, though its interpretation can become more dramatic than the source material warrants.
Blender MCP and Mixamo make conversational 3D creation and motion-capture workflows much more accessible.
Wolfram Alpha remains a useful structured-knowledge layer for educational products where predictable sourcing and school-safe answers matter.
Platforms are pushing back on AI-only media, which increases the value of tools that amplify authentic expertise instead of manufacturing a synthetic personality.

The hard part is no longer capability

The projects at this meetup did not share a stack or a business model. They shared an operating problem.

AI systems can write code, render video, organize relationships, build 3D scenes, synthesize audio, and operate workflows. But the more capable they become, the more important the surrounding systems become: schedules, budgets, review loops, privacy boundaries, abuse controls, and interfaces that keep human judgment in the right places.

That is where the most interesting product work is happening now. Not “can the model do this?” but “can this become a dependable part of someone’s day?”