Shape Rotator · What did you get done last week · NY · 8 June 2026

Building stopped being
the hard part.

The cohort's first standup at the Shape Rotator accelerator — and the gap between each project's one-line pitch and where its hunt for product–market fit actually stands.

Co-facilitated in the room, with elizaOS joining remotely and a visiting mentor present for the first half. The cohort gathered in person at The Convent. Shape Rotator accelerator — run by Flashbots X and IC3, with support from The Convent.
~20 min read  ·  A site-vs-room Δ for every project, walked member by member, plus the themes that ran through the room.
The 60-second version

ContextThe room, and a note on the transcript

Shape Rotator is a ten-week NYC accelerator (May 18–July 25, 2026) that turns cryptography and AI research into products, run by Flashbots X and IC3 with support from The Convent. This was the cohort's first WDYDLW standup, held at The Convent. elizaOS — an open-source agentic operating system — co-facilitated remotely, framing the format as a riff on the “what did you get done this week” email: honest accountability mixed with a short demo. A facilitator ran the room, with a visiting mentor there for the first half; the rest of the cohort joined in person.

Why attributions are reconstructed
Most of the cohort was together in the room, captured through a single in-room device, so the auto-transcript collapsed nearly every in-person voice — the facilitation and each member's update alike — under one label, while the remote participant came through cleanly. Who-said-what here is reconstructed by matching each update to its project on the cohort roster; individual lines are attributed at the project level, and uncertain names are flagged in the notes.
Advanced — moved toward its milestone Reframed — pivoted the wedge or ideal customer profile (ICP) Hard truth — a gap surfaced stage 1–7 = the cohort's own PMF-journey scale (1 = idea, 7 = scaling).

Around the roomWho shipped what — member by member

Each builder's update, grouped by person (one person may carry several projects). For each project: the pitch registered on shaperotator.xyz, what the room reported, and the Δ on its hunt for product–market fit.

Elocute

ElocuteIn the room

Elocute

Consumer · speaking practicestage 5/7Retention
On the site

Live AI speaking-practice app (elocute.fun): 538 users, ~39% 30-day active, strong organic content reach. Next milestone: turn active users into a repeatable weekly practice loop with a clearer paid-conversion signal.

In the room

The team ran its first structured user-research interview (with the TeeSQL team), prompted by cohort feedback, and designed three research cycles — interview three users, ship updates, then re-interview those three plus three new ones. The work is compressing a flood of conversations into one executable product/UX plan ahead of the local hackathon; elizaOS pitched in on shipping recent app improvements.

Δ PMF · Advanced

Attacking the retention bottleneck with qualitative discovery rather than more features — “one real user interview beat five days of speculation.” The update was about turning feedback into a concrete product plan, not a form-factor change.

elizaOS

Co-facilitatorRemote

elizaOS

AI · agent OSstage 7/7GTM
On the site

Mature open-source agentic operating system (runtime, cloud, desktop, mobile, plugins). The open question isn't validation — it's GTM. Next milestone: turn the broad ecosystem into one or two concrete cohort integrations with real downstream pull.

In the room

Hard pivot to a new agent interface (the team expected a rival, Hermes, to ship a desktop app, so Eliza reworked its own). Swapped OpenRouter → cohort-mate Bitrouter (OpenRouter kept as fallback); swapped n8n → cohort-mate Smithers as the workflow engine, shipping PRs back to it (GEPA reflective prompt self-optimization, a TUI). Built a sandboxed tiny-agent on Tinycloud; stood up hosted Eliza agents (Eliza Cloud); and ran a side stunt using agents to resolve most of the open lemmas in the arkworks zkSNARK library — by the team's telling, years of proof work compressed into days, though the top prize stayed out of reach.

Δ PMF · Reframed

The integration milestone the site asked for effectively landed — Eliza now runs on two cohort tools. the team named the real constraint out loud: “we don't have a product-creation problem, we have a product-sales problem.” Focus moved from capability to a buyer definition and last-mile distribution; an MVP (cloud, multiplayer, power users) and a possible non-technical co-founder are now on the table.

TeeSQL

TeeSQLIn the room

TeeSQL

Infra · confidential DBstage 4/7Technical Risk
On the site

TEE Postgres on dstack, with an RA-TLS proxy and open-source attestation tooling. Next milestone: onboard one cohort team to the beta and document the full attested connection path.

In the room

Product-shape pivot: from a Postgres-specific high-availability cluster to a generalized, attestation-gated mesh that can run any open-source software (Clickhouse, Redis), with a blockchain control plane and host- or dev-proof modes. The team is hunting an ideal customer profile (ICP) off-cohort — European confidential-computing enterprises (e.g., a German firm reducing US-cloud reliance) reached via a web2 enterprise network.

Δ PMF · Reframed

Broadened the surface from one database to any confidential workload, and moved the customer search toward web2 enterprise. Candid self-assessment: “nothing fundamentally defensible yet.” The bottleneck is migrating from technical risk toward ICP clarity and a moat.

Teleport Router & Feedling

Co-facilitator (first half)In the room

Teleport Router

Protocol · routingstage 3/7Solution Quality
On the site

Cross-network routing; public surface is one-time-use delegated posting (teleport.best). Next milestone: land one real cohort use case where another project routes an action through Teleport.

In the room

A new Daybook electron app and the “Day” project — a local daily log of your Claude-Code/Codex sessions — with a cloud-SDK onboarding flow; demoed at the Princeton event. The Teleport Router was PR'd into Shape Rotator OS — the first merged cohort-project-to-Shape-Rotator-OS contribution.

Δ PMF · Advanced

The “route through Teleport for a real cohort use case” milestone began to land via the Shape OS merge. The product is drifting from delegated posting toward a local-first daily log and a cohort cross-pollination feed.

IC3-camp hackathon project

Side project · hackathon

The IC3-camp hackathon side project — a “try-before-you-buy” marketplace for proprietary agent skills, where a buyer benchmarks a skill against their own withheld test set and, once it clears the bar, licenses it to run inside a TEE so the skill stays private — won the camp's “funniest project” award.

Feedling

Consumer · feed / companionstage 2/7ICP Clarity
On the site

Information-diet interventions for short-form video, using TEE-delegated watch history. Next milestone: create the first live user thread and define one quickly-testable intervention.

In the room

An update for the Feedling team: a second consumer app — a customizable AI companion — is showing early signs of interest from the Xiaohongshu (RED) community.

Δ PMF · Reframed

An expansion from the site's “info-diet for short-form video” toward a live AI-companion app with early community pull — alongside Elocute, one of the room's clearer consumer-traction signals. The PMF question shifts to retention and onboarding.

DarkBloom security audit

Personal update · security audit

Separately, a personal security audit of EigenLayer's DarkBloom was shared — the project turning idle Macs into a decentralized inference network. The read: its Mac remote-attestation doesn't deliver the security it claims. Unrelated to Feedling; it ties back to the cohort's Agentic Organizations salon.

Daedalus

DaedalusIn the room

Daedalus

Crypto · prediction marketsstage 2/7ICP Clarity
On the site

Prediction-market microstructure research grounded in the PROF paper; building an orderbook-ingestion + backtest pipeline. Next milestone: show one backtest or live result that changes a market-making decision.

In the room

Product = using Polymarket + sports events to hedge stock portfolios. The team cited Polymarket's first institutional block trade (FalconX, an AI-compute price index) as evidence of institutionalization, and is scraping Polymarket historical orderbook data (L2/L3) to quantify it; a new demo lands by end of next week.

Δ PMF · Reframed

A facilitator's heavy steer de-risked the product from a “make-users-money” bot (fiduciary / broker-dealer exposure) toward a discovery tool for delta-neutral hedges — gated on rigorous correlation backtesting before any app is built. ICP sharpening on the institutionalization signal; the core correlation is still unproven.

Tinycloud

Tinycloud (team)In the room

Tinycloud

Infra · user-owned cloudstage 4/7ICP Clarity
On the site

User-owned cloud on dstack TEEs; live SDK + protocol; paid pilot with Sparq Gaming. Next milestone: get one cohort project using tinycloud-secrets for a real scoped-delegation workflow.

In the room

Shipped single-message delegation in the encryption/sign protocol — a major simplification over interactive key exchange. tinycloud-secrets now powers “Listen,” a transcript aggregator (Firefly/Granola/…). Shipped Tiny Cloud Chats (built on RedPill) for private AI chat, and finalized a policy engine that gates content by credential and trust graph. Building a personalized feed that mines a user's own transcripts into artifacts.

Δ PMF · Advanced

The scoped-delegation milestone effectively met, with a live internal consumer (Listen). Surface expanded toward a consumer feed + policy engine; the customer search is sharpening around “people with lots of private transcripts,” with a clear contrast drawn vs. Teleport (personal aggregation vs. cross-pollination).

Contexto

ContextoIn the room

Contexto

AI · context enginestage 3/7Solution Quality
On the site

Agent context engine; episode-based memory across runtimes (OpenClaw plugin). Next milestone: instrument one cohort agent workflow end-to-end and show better recovery/continuity.

In the room

A deep dive on coordination topologies — Paradigm's “Centaur” — and a design where user preferences and code logs stay on local edge devices while routing happens through a centralized, privacy-respecting router. Built a personal to-do/journaling agent on Hermes; interviewed executive assistants on LinkedIn (a $200–$5,000/mo personal- and executive-assistant market).

Δ PMF · Reframed

A facilitator's hard steer reframes the Contexto wedge from agent-coordination infra toward a mass-market personal assistant — “a market 1,000–1,000,000× bigger.” Another facilitator pushed back: agentic executive assistance is a brutally hard product — too personal to get right, with impatient, error-intolerant buyers. The wedge is contested; the near-term move is a personal-assistant PoC plus more discovery.

JJHub · Smithers

JJHub · SmithersIn the room

JJHub / Smithers

AI · agentic codingstage 4/7ICP Clarity
On the site

Agentic coding platform; Smithers = durable agentic workflows as JSX (hit the HN front page); Tevm has an EF grant. Next milestone: get one cohort team using Smithers for a real agentic CI or review loop.

In the room

Shipped the long-awaited UI (“the dots are actually good now”) — pixel-polishing the single most-requested feature for ~3–4 months. Recorded a podcast with an Ethereum Foundation researcher. Reframed positioning: “Smithers could be the Linear of workflow tools” (vs. Zapier / monday.com). The product stabilized from maintainer-dependent to genuinely usable; ~6 organic onboards, roughly doubling week-over-week with no marketing.

Δ PMF · Advanced

The site's cohort-integration milestone is landing through Eliza (elizaOS adopted Smithers as its workflow engine). The product crossed from fragile to stable → organic pull. PMF hunt now centers on serving a non-self ICP and a clean one-line position; a non-technical co-founder is being floated.

DealProof

DealProofIn the room

DealProof

Protocol · agent contractsstage 4/7Technical Risk
On the site

TEE + post-quantum dual-agent contracts (Intel DCAP, on-chain escrow); 56 passing tests; Bradford Quantum Hackathon winner. Next milestone: get one cohort team to integrate the dual-attestation contract layer.

In the room

Enhanced the agent-to-agent negotiation app using the Contexto memory engine — CDN deal negotiations now carry memory across rounds (store a hash, recall prior terms). Implemented PiCred (arXiv:2606.03771) for provenance — verifying an agent's code wasn't tampered with — and is introducing the paper's authors to the cohort. Began scoping a product (an API platform inside a CVM) and cold-reached two AI-negotiation companies (e.g., pactum.ai) to find the pain.

Δ PMF · Advanced

From a dual-attestation primitive toward a concrete negotiation product, a provenance integration (PiCred), and first customer-discovery outreach. Now actively cross-pollinating with Contexto (memory) and Tinycloud (provenance).

Bitrouter

BitrouterIn the room

Bitrouter

Infra · LLM routerstage 4/7ICP Clarity
On the site

P2P LLM router; live at bitrouter.ai; opening up x402-kit and accepting cohort dogfooders; Phala-CTO referral. Next milestone: get a cohort team using x402-kit / Bitrouter inside a real agent workflow.

In the room

Repositioned the site from “open intelligence router” to a sharp pain point: reliability + cost for coding agents. After a facilitator retweeted it, GitHub stars doubled. Shipped a feature combining Claude/Codex subscriptions with open-source models for cheaper workflows. Wired in analytics: ~10% of signups activate; zero have paid — now reaching out to non-payers one by one. Supply-side edge: legal, cheap Chinese tokens via the Phala network.

Δ PMF · Hard truth

Sharper positioning plus a real distribution spike (stars 2×) — but the conversion truth surfaced: 0 paid. The PMF hunt pivoted to a supply-side moat (token deals) and a conversion funnel. As one facilitator put it, it's a red-ocean aggregator game where the edge is cost and grind, not features.

Shape Rotator OS & Private search

Shape Rotator OS · Private LLM searchIn the room

Shape Rotator OS

Infra · coordinationstage 2/7Product-driven dev
On the site

The cohort's own coordination layer and cohort viewer (dmarzzz/shape-rotator-os). Next milestone: open it to outside contributions and real cohort workflows.

In the room

Re-architected Shape OS for outside contributions and flagged the need for real user stories and product-driven development. The Teleport Router PR was the first merged cohort-project-to-Shape-Rotator-OS contribution.

Δ PMF · Advanced

Coordination OS opened to external contributors (first project merged in). Honest read: no PMF, but a working example of cohort coordination.

Private LLM search

Infra · private searchstage 2/7Technical Risk
On the site

searxng-wth-frnds — a LAN-first peer search daemon — plus DCNet-style anonymous broadcast. Next milestone: one real cohort workflow using private search, and a minimal anonymity architecture note.

In the room

Private LLM search with metadata privacy via flashnet. Tor exit nodes get blocked fast (Cloudflare/AI bans), so the work explored reputation-gated egress to prevent IP poisoning; a deployed prototype benchmarked 10k queries at median +1.2s latency, p99 ~9s.

Δ PMF · Advanced

Moved from idea → deployed, benchmarked prototype with a clear latency envelope (fine for agentic search, not low-latency). Metadata privacy now routed through flashnet.

PatternsThemes that ran through the room

The same handful of ideas surfaced in nearly every update.

01

Code is commoditized; selling isn't

If an MVP of almost anything can be built in two weeks, the table-stakes part is done. The hard, unsolved part is naming a customer who will actually pay — and most teams admitted they were still guessing at theirs.

02

Don't be a fiduciary

For the finance-adjacent teams, a facilitator drew a bright line: building a bot that makes other people money is a fast path to fiduciary and broker-dealer risk. Build discovery and intelligence tools instead — and back-test the correlation before building the app.

03

Reach the last mile; sell to normies

Developers and open-source maintainers are the worst customers to sell to. The far larger market is ordinary users who want an agent that just works — so own the end product rather than depending on platforms that can absorb you.

04

Aim for a six-month moat

Nobody expects a permanent edge when everything is one-shottable. The realistic goal is a defensible six months — bought with cheap supply, private data, real relationships, and speed.

05

The cohort is the distribution

Eliza now runs on Bitrouter and Smithers; Teleport's router merged into Shape OS; DealProof is wiring in Contexto and py-credits. A facilitator's pitch for ETH Global: bundle these into one demo and share attribution — collaborative distribution over solo launches.

“Just assume everything can be one-shotted — then build from that world.”
A facilitator, on where the real work now lives

The weekOne product, one hundred messages

The cohort set a single forcing function for the midpoint week: each team ships a concrete, usable product to demo at the ETH Global New York hackathon (turn-in Sunday), and each founder sends 100 cold messages to their ideal users to validate direction. Teams that can't field a full crew were urged to bundle into a shared project and split attribution, and to step up project managers since several members travel that week. The accelerator also floated a half-serious metric — the “one-shot challenge”: if the facilitators' agents can one-shot 80% of your product, that's your cue to question the moat and find the niche only you can serve.

Cross-reference
The DarkBloom audit above examines the same Apple-Silicon inference project Eigen Labs presented at the cohort's Agentic Organizations salon. The read: its Mac remote-attestation doesn't deliver the security it claims.

ReferenceGlossary

WDYDLW
“What did you get done last week” — the recurring accountability standup format.
PMF
Product–market fit: evidence that a defined customer actually wants and will pay for the product.
ICP
Ideal customer profile — the specific buyer a team is targeting.
TEE / CVM
Trusted Execution Environment / Confidential VM: hardware-isolated compute that even the operator can't observe or tamper with.
Remote attestation / RA-TLS
Cryptographic proof that specific code is running inside a genuine TEE; RA-TLS binds that proof to a network connection.
dstack
An open framework for deploying confidential containers in TEEs, used across several cohort projects.
LLM router
Middleware that sends each request to the cheapest/best model; OpenRouter is the incumbent, Bitrouter the cohort entrant.
Smithers
JJHub's tool for durable agentic workflows written as JSX — an open-source take on workflow automation.
Scoped delegation
Granting an agent narrow, revocable access to a credential without handing over the whole account (Tinycloud).
PiCred
A research method (arXiv:2606.03771) for proving an agent's code wasn't tampered with — provenance for agents; integrated by DealProof.
Delta-neutral hedge
A position built so the trader doesn't lose regardless of market direction — the discovery target for the prediction-market team.
Centaur
A coordination topology (from Paradigm) for human+agent teams that Contexto studied for routing design.
One-shot
To recreate a product in a single agent pass; the bar against which moats are now measured.
x402-kit
Bitrouter's modular SDK for HTTP-native (x402) agent payments.
DarkBloom
Eigen Labs' project turning idle Macs into a decentralized inference network — audited here for attestation flaws.
RedPill
The private-AI-chat layer Tinycloud built Tiny Cloud Chats on.
flashnet
The metadata-privacy network the Shape OS team routes private LLM search through.
GEPA
A reflective DSPy prompt optimizer (Genetic-Pareto); shipped into Smithers by elizaOS.
arkworks
The Rust zkSNARK library ecosystem whose proofs elizaOS's agents worked through.

AppendixBaselines & sources

Each project's “On the site” lane is drawn from the cohort's own public profiles — the same data that powers the cohort viewer at shaperotator.xyz / os-web.shaperotator.xyz — including each team's registered focus, ICP, journey stage, primary bottleneck, and next milestone. The “In the room” lane and Δ are reconstructed from the 8 June standup transcript.

Externally verified

Notes

  1. Speaker attribution is reconstructed. The cohort was together in the room at The Convent on a single capture device, so the transcript labeled most in-person voices identically (including the facilitation and each member's update); the remote participant was captured separately. Updates are matched to people and projects by content and the cohort roster.
  2. The “funniest project” try-before-you-buy skills demo is an IC3-camp hackathon side project, not a separate cohort team's update.
  3. elizaOS terms were corrected against public sources: the audio's “Japa” is GEPA (a reflective DSPy prompt optimizer), and the proof library is arkworks (the Rust zkSNARK ecosystem). The “$1M grand prize” is the team's framing; the specific prize could not be confirmed.
  4. DealProof's provenance method is PiCred (arXiv:2606.03771).
  5. Journey stages, bottlenecks, customer profiles and milestones in the “On the site” lane are the teams' own self-reported values from the cohort dataset, not an external assessment.