AI Business Builders / Atlas methods

Twelve selected stories: source notes

Back to the Atlas | Methods

The text below preserves the corrected snapshot. These are attributed demonstrations and advertised formats, not independently reproduced outcomes. Invitation covers are not event photographs.

A fly-brain model flying a drone

Evidence: Builder-reported demo

Builder Matthew Hoe reported running a FlyWire connectome model of about 140,000 neurons on a GPU to control a small drone, with an avoidance response when a hand approached.

Biological circuit structure as a control experiment. The implementation and performance have not been independently verified here.

Visual: Video preview from the builder's X post. Retrieved 2026-10-17. Media source.

AI apps built to work offline

Evidence: Listing + organizer recap

Zetic and Liquid AI asked teams to build phone-based assistants for an imaginary lunar mission. Zetic later reported that the apps kept working offline, including a translator and spacecraft guide.

A concrete constraint for a demo: make it work on the device, with the network disabled. The moon was the workshop scenario, not a deployment.

Visual: Event photo from the linked X recap. Retrieved 2026-10-17. Media source.

Voice agents tested with live phone calls

Evidence: Listing + organizer recap

Cresta reports that 30 builders made 10 voice agents in under three hours. Judges then challenged the agents with unscripted live calls at the Anthropic-sponsored hackathon.

Evaluation was part of the show: the agents had to handle questions the builders had not scripted.

Visual: Event photo from the linked X recap. Retrieved 2026-10-17. Media source.

AI security testing with replayable proofs

Evidence: Listing + challenge recap

Semgrep describes a winning Reva project that combined static checks with model reasoning, tested suspected issues in a sandbox, and checked proposed fixes again.

The interesting design choice is verification between model suspicion and a developer-facing report. This is a reported hackathon design, not a production benchmark.

Visual: Architecture diagram from the builder's X post. Retrieved 2026-10-17. Media source.

Ordering matcha through an AI agent

Evidence: Listing + participant recap

The Agent Matcha Popup offered a drink ordered through Monid. Grace Ling reported connecting Monid to Claude Cowork, choosing the milk and having the agent place her order.

A small, tangible agent workflow: a preference became an order through a connected service.

Visual: Video preview from a participant's X post. Retrieved 2026-10-17. Media source.

An AI agent in a fictional heist

Evidence: Listing + host recap

Cogent, Modal and LangChain advertised a timed challenge in an isolated cloud environment: direct an agent through staged security tasks to find a proof token. A host later posted a recap.

A bounded game turns tool use, debugging and security reasoning into something people can watch and compare.

Visual: Event photo from the linked X recap. Retrieved 2026-10-17. Media source.

A manicure with an NFC chip

Evidence: Archived program listing

Cyberglam advertised NTAG215 nail-art stickers and a Flipper Zero for programming them. The listing imagined a business card or key fob embedded in a nail.

A wearable interface in a tiny form factor. The archived listing does not establish which access systems worked or confirm the results.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.

Karaoke with AI-generated songs

Evidence: Archived program listing

Andala Labs invited guests to submit a prompt, have a song generated and sing it. Suggested genres included sea shanties about seed rounds and Gregorian chants about late-night coding.

The product demonstration was a participatory performance. The source is the event description, not a verified report of the performances.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.

Startup pitches roasted by a drag queen

Evidence: Archived program listing

StartOut advertised a pitch night with Elsa Touche roasting six founders, followed by investor questions and audience voting. Each pitch had a three-minute slot.

Comedy and investor feedback were deliberately put on the same stage. The format comes from the listing.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.

A potato score for your startup brand

Evidence: Archived program listing

Tonik invited startups to submit websites for live feedback from Lester Chen, Morgane Palomares and Malthe Sigurdsson, with Maya Spivak leading the conversation. The audience would vote with potato paddles.

Potato Meter was a playful brand-review format, not a hardware device or a new measurement algorithm.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.

Coffee, cockroach racing and RNA

Evidence: Archived program listing

Yeast Bay Bio and California Innovation Fund advertised a morning combining RNAi biopesticide discussion, cockroach races and photos with stick insects.

An unusual way to introduce a biotech product. Safety and efficacy language in the invitation is a company claim, not a finding of this Atlas.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.

An escape room followed by an AI hackathon

Evidence: Archived program listing

The CTRL hackathon started in one of three escape rooms. Teams were then meant to unlock a mission, receive Devin credits and build with the agents of their choice.

The physical puzzle supplied the story and the software brief. This is the advertised format; a completion recap has not been matched.

Visual: Archived invitation artwork. Retrieved 2026-10-14. Media source.