Developer Projects

**Project description** Popper is an adversarial verification gate for pull requests. We built it after AI coding agents began producing fixes faster than we could confidently review them. A green test was not always proof—it sometimes passed before the fix too. Popper extracts the behavioral claim behind a pull request, generates tests designed to break that claim, and executes each test against both versions of the code: **Fail before + Pass after = Evidence of a fix** Fireworks extracts the claim and generates adversarial tests. Daytona runs them safely in isolated sandboxes. CodeRabbit provides an independent static review, which Popper compares with the executed evidence while keeping opinion and proof clearly separated. Braintrust traces the pipeline, and CopilotKit lets reviewers ask questions about the results. Popper flags tests that pass on both versions as inconclusive and treats sandbox failures as missing evidence—not failed code. It then presents the claim, test results, disagreements, and a recommendation. A human always makes the final merge decision. We built Popper with Next.js and TypeScript, plus a replay system that can instantly load a previously verified run if a live service becomes unavailable.

Every canceled appointment is a perishable asset. Blankless is an autonomous appointment-recovery system for outpatient clinics, starting with dermatology. When a cancellation opens a slot, Blankless immediately filters the waitlist, blocks ineligible candidates using named policy rules, ranks the remaining patients, and conducts sequential outreach until the appointment is filled. The workflow is transparent rather than a black box. Every decision is recorded in an audit timeline, including why a patient was excluded, how candidates were scored, who was contacted, and the final outcome. The system then learns from that outcome through a versioned policy. In the demo, Sofia declines an afternoon appointment, Maya accepts, and Blankless updates its time-of-day preference weight. In the second scenario, the learned policy ranks Chloe first and fills the next opening in one attempt. Blankless reduces front-desk work, recovers otherwise lost revenue, and gives patients faster access to care. The core is a deterministic Next.js and TypeScript workflow engine built for reliability, testing, and future integration with scheduling systems, messaging platforms, and human escalation.

EduForge is a living classroom that rebuilds itself in real time as an agent mesh works. A professor uploads an assignment; eight typed agents extract its concepts, then group students into temporary rooms (Ember, Forge, Harbor, Summit) by their current learning barrier, never by diagnosis or label. Each room gets an assignment variant that keeps the objective and rigor but changes the pathway. Submissions run, assessment surfaces misconceptions, mastery updates, students move, rooms resize, and a next-day plan appears with evidence for every decision. Every backend event drives an animation in a Phaser pixel world, so you watch the school reorganize itself. We used five sponsor tools: Guild.ai is our control plane, it registers every agent under least-privilege permissions, gates low-confidence grades and the final plan for human approval, and audits every action. Band is the communication mesh, it carries each agent decision as a typed message and turns it into an on-screen animation, with full history for replay. Actian is the classroom memory, it stores student-by-concept context, retrieves it by relevance, and keeps evidence-linked observations and room snapshots. Pioneer is the reasoning engine, concept extraction, variant generation, misconception explanations, and lesson synthesis, every output Zod-validated. Replay.io records the session with a marker at each demo beat, so any run stays debuggable. Built with Next.js, TypeScript, Phaser, Zod, and SSE.

Galya's World Taste Model can predict a user's taste, we use it along with Band and Pioneer to run a simulation to demonstrate a self improving agent where the agent learns a user's conversational and problem solving taste. For this hackathon specifically we trained a specific version of our model on LLM conversational datasets. We wanted to also train it on images and show how the conversational user history can be transposed in our model to predict visual taste as well. We dropped this for lack of time. We rebuilt the agent using Guild as well.

Invisible Hand is a running market, not a chatbot demo. A population of seller-agents sells claim extraction over live Hacker News articles behind x402 paywalls. Buyer-agents pay real testnet USDC on Base Sepolia (250+ settlements with tx hashes), Gemini adversarially verifies every claim against the live source, and a verified failure sharply cuts a seller's future demand. Cumulative net P&L is the fitness function: one variant went insolvent and its endpoint now literally returns HTTP 410 GONE; the strongest variants breed through a six-rule fail-closed governance gate, with children inheriting immunity from their parents' failure clusters stored in Actian VectorAI. During the recorded demo the market bred a new variant through that gate live on camera. It also caught its own first case of reward hacking: generation-1 economics improved several-fold while verified accuracy declined, and the dashboard reports that honestly (receipt-true successes/attempts). Pioneer serves the inference being sold; model routing is a gene. BAND carries a real two-agent exchange: the buyer @mentions the market herald and the herald answers each settlement. Senso ingests every generation report and its engine published a public cited.md article from them. Replay QA ran two rounds; all four defects found were fixed with regression tests. Every non-live integration is labeled local mode; everything on-chain is independently verifiable. Demo Video: https://youtu.be/XA3-3MLTkAM

LivingBook — books that talk back. Live: livingbook.masky.ai (web + iOS). Real authors' books become talking video agents that answer ONLY in the author's exact words — ask the Bible about today's headline and a narrator reads you the verse on video. Authors sign in with Masky (avatar = identity), claim books, upload podcasts/fan Q&A, or record spoken sessions gated by VoiceCert human verification, so agents can say "I spoke these words myself — attested." Readers get byte-exact quotes with provenance; a News tab gives each book's grounded take on live headlines; Talk lets you converse as your avatar (lands in the author's inbox, rendered as two-avatar video on their credits); a podcast generator stages your books discussing a story you pick. Self-evolving 3 ways: the KB grows from use (every Q&A + attested session re-ingested); the trend-watching marketing agent ships versioned strategy on Guild.ai with rollback + audit (v1→v2 live today); Replay QA simulates readers and feeds back fixes. Sponsors, all live: Actian VectorAI (verbatim KB, Docker/edge), Pioneer (grounded reasoning under a no-quote contract), Guild.ai (governed runtime), Band (real agent-to-agent podcast rooms), Senso (verified ingestion + GEO), VoiceCert (human proof), Masky (SSO + all video), AWS. Enforced, not promised: LLMs select and reason; byte-identical checks reject invented quotes. The repo ships its full human/AI session log (Open Session License).

Imported from GitHub for a live builder workflow test. Repomix turns a repository into an AI-readable package for LLM review, context sharing, and coding-agent workflows. Imported signals: README found, package.json found, repo URL validated, demo URL detected. What I would publish as a developer: a concise project proof page with repo, demo, stack notes, and launch context. The raw README import was cleaned up before publishing.