Episode 9
This episode is almost entirely a product teardown. On this episode we bring on Josh Guter, founder of Elephants, for a live demo of the AI assistant running his ops platform, and then spend the back half grilling him on the stuff every founder building an AI chat feature actually loses sleep over: which model, what it can touch, how you stop it from doing something dumb, and how you audit a system that's too complex to fully read anymore.
Who the Guest Is
Josh Guter is the founder of Elephants, operations software built for small teams (think one to five people) who are managing big networks of relationships and projects without the bandwidth to set up (or maintain) a proper Notion or Airtable system. Elephants ships ready to go out of the box, with an AI assistant at the center that acts as a thought partner and de facto ops person for teams that can't afford to hire one. Josh studied graphic communications, not computer science, and taught himself to write software along the way. His first startup built embedded firmware for wearable devices that let hikers in remote areas signal for help with the press of a button. He's also the program manager for NC State's Entrepreneurship Clinic, which has been running on Elephants for about six months.
What We Cover
64 tools, one job each: Elephants’ assistant works through roughly 64 discrete backend tools: search relationships, search projects, stage a card, and so on. Each has tight instructions for when the model should reach for it. It’s the same architecture Josh first built as Make.com scenarios wired to OpenAI, just ported into full-stack code.
Nothing writes without a click: By default, the assistant can’t touch the database directly. Ask it to “create 10 catering tasks” and it comes back as a batch of cards you have to review and hit “add” on. Even in auto mode, deletions and archives always require human approval. No exceptions.
Breadcrumbs instead of a black box: Every tool call, error, and decision path gets logged to a table Josh can pull up when a customer says the assistant did something weird. It’s his stand-in for the old Make.com flow he used to watch run step by step in real time.
“That’s kind of scary”: Josh’s own words about how much of the current codebase has become genuinely opaque to him. His fix is opening a fresh Claude Code session and asking it to explain the system to him “like I’m 5” before he goes digging himself.
Illustrator mockup → Claude plan → Codex teardown: His build loop starts with sketching the UI in Adobe Illustrator, a holdover from his graphic communications degree. Then he has Claude draft an implementation plan and gets Codex to “tear it apart” for broken conventions and repeated logic before a line of code ships.
Red-teaming himself: Before any change touches the assistant’s tools or system prompt, Josh tries to break it in a dedicated “assistant stress test space,” including trying to get it to mass-delete every note in a workspace just to confirm it gets rejected.
From Glide + Make.com to full-stack: Elephants started as a low-code app on Glide Apps with an AI “brain” wired through Make.com automation scenarios. The full rebuild ran on Opus 4.8 with Codex reviewing, using the old Make scenarios as the literal spec for what to replicate.
Six months running a real entrepreneurial clinic: Elephants has managed NC State’s Entrepreneurship Clinic for about half a year. Josh’s proof point: he asked the assistant which program contacts to check in with ahead of a leadership transition, and it correctly surfaced four or five based on past conversation history.
A PWA on purpose: Elephants skipped the App Store entirely. It’s a progressive web app that works identically on desktop, mobile browser, or installed to a phone home screen.
The honest solo-founder moment: Josh admits to sometimes wondering if he’s “neck deep in AI psychosis” building something this consequential alone. He says hearing other founders describe the exact same workflow is what keeps him grounded.
Josh's closing line sums up why this one's worth a listen: the more teams trust these systems with their real operational knowledge, the more useful they get, and the harder it becomes to know exactly what's happening under the hood. Enjoy the conversation.
It’s best viewed on YouTube to fully see the examples (make sure to subscribe!)
But also available on all audio podcast players through Tweener Talks!
PLUS we have a new spot for show notes and files discussed in the episode. Check it out: https://github.com/instanttaylor/redacted-podcast
We need your [Redacted] AI experiences for upcoming episodes! Who’s the most underrated AI builder you know? Someone running real systems inside a real business? Send us a message at contact@tweenerfund.com because we want to get them on the show!
What’s Next?
New episodes drop twice a month/every other Wednesday. If you want to be on the show as a guest and show your [REDACTED] builds, email us here.