Event projects
Explore the work shared by participants in the August 8, 2026 Claude Impact Lab.
- #01via MCPCreative work under pressure
A tool that helps creatives unsure of AI context and knowledge to guide their career.
Ground turns six honest answers about a creative's actual practice into a personalized AI Adaptation Blueprint — a career plan with every claim traceable back to a specific answer, so nothing is invented or generic. Attached: the submission deck (PDF) covering the problem, the Run A/Run B proof, and the product.
GROUNDClaude APIground-ai-blueprint.netlify.app - #02formCreative work under pressure
A tool that helps creatives unsure of AI context and knowledge to guide their career.
What we built: Ground, a SaaS tool that turns six honest answers about a creative's actual practice into a personalized "AI Adaptation Blueprint" — a career plan grounded entirely in what they told it, with every claim traceable back to a specific answer (hover any line in the report, see the source). Built out as a working interactive prototype (deployed at ground-ai-blueprint.netlify.app), a 5-screen Figma product design, and validated with real QA: a side-by-side comparison showing generic AI output (no context) versus Ground's grounded output (six real answers) for an illustrator, copywriter, and musician. Problem it solves: Creatives are anxious about AI for two compounding reasons — they don't understand it, and most of what reaches them is generic "AI will change everything" content that could apply to anyone, which reads as slop and deepens distrust. Ground refuses to generate anything until the person brings real specifics, then proves — visibly, not just by policy — that nothing in their plan was invented or copy-pasted from a template. It's a forcing function against generic AI advice, not another AI-fear explainer.
GROUNDclaude APIground-ai-blueprint.netlify.app - #03via MCPCreative work under pressure
Rider — the starter pack for your working life in music
Music policy, funding and community are built for songwriters and artists. Engineers, session players, live crew and educators are invisible: a correct query against the federal nonprofit registry returns zero of the organisations they belong to. Rider takes what you do, in your own words, plus your city, and returns one concrete action today, backed by 8 live federal bills, 29 real grants and 40 verified orgs. Click any bill and it finds your senators with real phone numbers and drafts your message. Working demo, running locally.
Team Rough RidersAstro, Claude Code, congress.gov, congress-legislators, IRS Exempt Organizations filerider-music.pages.dev - #05formWildcard
Right Knock
We built Right Knock — point your phone at a problem, and we tell you the one government office that can actually fix it, with the deadline, the form, and a ready-to-send script. Here's the problem it solves. Everything wrong in your life — a landlord stiffing you, a boss underpaying you, a streetlight that's been dead for months — has exactly one office with the power to fix it. And almost nobody can find it. Government is thousands of doors: city, county, state, federal, special districts. The boundaries are invisible, and your mailing address lies — millions of people 'live' in cities they don't legally live in. Stand on one corner in LA and the minimum wage is $17.87. Walk ten minutes and it's $20.22, with a different enforcement office and a different deadline. Nobody knows this. So people get bounced between offices until the deadline passes — and when the deadline passes, the right dies with it. The wrong door isn't an inconvenience. It's how people lose. Right Knock fixes the step every other tool skips. You give us your location, a photo, or the notice on your door — in any language. We resolve who actually governs that exact spot for that exact problem, then hand you an action, not an answer: the agency, the rule, the deadline, the evidence to gather, and a script ready to send — translated. 311 only works where the city bought it. 211 is a phone book. ChatGPT is confidently wrong at the block level. We're the only ones whose core question is the one everyone else assumes away: who's actually in charge, right here? Let me show you. This streetlight has been out for months..
Right KnockClaude Api, Convex, Next, React, Clerk, MCPgithub.com - #06formWildcard
KneadIt
KneadIt solves this problem by giving local businesses a way to sell their day old bakery items instead of throwing it away to people looking for affordable food.
KneadIthttps - #07via MCPCreative work under pressure
Rider — the starter pack for your working life in music
Music policy, funding and community are built for songwriters and artists. Engineers, session players, live crew and educators are invisible: a correct query against the federal nonprofit registry returns zero of the organisations they belong to. Rider takes what you do, in your own words, plus your city, and returns one concrete action today, backed by 8 live federal bills, 29 real grants and 40 verified orgs. Click any bill and it finds your senators with real phone numbers and drafts your message. Working demo, running locally.
Team Rough RidersAstro, Claude Code, congress.gov, congress-legislators, IRS Exempt Organizations file - #08via MCPWildcard
Navi
This app brings together public data sources and community resources to help individuals demystify the process of finding and applying for health insurance coverage. The goal is to remove the guesswork from a confusing system by giving users a clear, personalized answer
JACSClaude API (claude-opus-5) with streaming and tool use, in-process MCP server via @modelcontextprotocol/sdk, Bun (Bun.serve), React 19, Tailwind CSS 4, shadcn/ui, docx for Word generationgithub.com - #09formCreative work under pressure
A tool that catches burnout while it's still fixable
We built a recovery-deficit detector for office workers. It reads wearable physiology, workspace environment, and two short daily check-ins, then answers one question each morning: is this person still recovering, or accumulating debt? When enough signals breach their personal baseline, Claude drafts a recommendation to the worker. It helps employees catch burnout before it costs them, and gives employers something measurable to fix. The core problem is timing. Companies spot burnout at the sick leave or the resignation, when it's most expensive and least fixable. We flag it days earlier by catching what self-report misses. Presentation: https://docs.google.com/presentation/d/1-7jzsMYI4P67OEjVpCtzTiESXpRhcZZTvgdVECuLGUM/edit?slide=id.g3f992e67e79_1_177#slide=id.g3f992e67e79_1_177
Team BeachdayClaude Opus, React, Elevenlabsgithub.com - #10formWildcard
Right Knock
LA County is one county, 88 separate cities, and about a million people living in unincorporated areas where no city government exists at all. Mailing addresses actively lie about this: an envelope that reads "Los Angeles, CA 90022" is usually unincorporated county, where City of LA agencies have no authority and often never tell you why they can't help. People lose wage claims and eviction deadlines to that confusion. rightknock.com
Right KnockClaude Api, Convex, Next, React, Clerk, MCPgithub.com - #11via MCPWildcard
Santa Monica Rent Control Rubber Duck
A citation-first AI agent for the Santa Monica rent control office, neutral between tenants and landlords. Residents wait a reported 3 to 5 weeks for an answer, and a generic chatbot will confidently invent the number. Here the self-check is real code, not a second AI opinion: every draft is verified against a fixed table of cited facts, and anything unverifiable is refused and escalated to a human rather than shown. Working zero-network demo in a single HTML file; the fact table is hand-maintained and the staff queue is honest scaffolding, not claimed as finished.
The Rubber DuckClaude, vanilla HTML/CSS/JS (zero dependencies, zero network, opens from file://), Python reference implementation of the critic loopgithub.com - #12via MCP
Navi
This app brings together public data sources and community resources to help individuals demystify the process of finding and applying for health insurance coverage. The goal is to remove the guesswork from a confusing system by giving users a clear, personalized answer
JACS - #13formWildcard
Check Your Politician
Check Your Politician Follow the money. Follow the votes.
Team PolitisourceClaude, Githubgithub.com - #14formCreative work under pressure
ShieldGuard — the anti-scam layer that fights back, without disclosing your identity.
The problem Scam calls cost Americans ~$25B+ a year, and the defenses all fail the same way. Carrier blocking is per-carrier and per-OS, so switching phones loses it. Blocking is passive: the scammer just calls the next person, and you have nothing to show for it. Reporting to the FTC produces no outcome you ever see. And the one real remedy — the TCPA's private right of action, $500 to $1,500 per call — is unusable, because collecting it means hiring a lawyer and handing your name, address, and phone number to the person who was just trying to rob you. So the rational move is to hang up. Which is exactly why it keeps working. What we built Four capabilities on one spine, tied to your digital ID rather than your carrier or your handset. Screen. An AI agent answers unknown calls, talks to the caller, and decides. It announces it's automated and recording (California is two-party consent — that disclosure is hardcoded, not model-generated). It never confirms a single detail about you and never absorbs urgency. Capture. The conversation becomes a SHA-256 hash-chained evidence record and clusters into a shared campaign database — 5-word shingles, Jaccard matching. One report is reported and never leaves the building. A second independent reporter promotes it to exportable. That filter runs in the query, not the UI, so there is no way to publish an uncorroborated accusation by rendering a different component. Enforce. Evidence is matched against a closed corpus of TCPA, FTC TSR, CAN-SPAM, and California rules. The model decides which statutes apply and why; the arithmetic runs in TypeScript. Findings split by whether you can actually sue: TCPA §227(b) and §227(c) go into your demand letter, TSR and CAN-SPAM go to the FTC and state AG, because you have no standing to demand money for those. Prevent. California DELETE Act deletion requests shrink the attack surface. The broker compliance deadline hit August 1 — eight days ago. The part that matters The letter agent is architecturally incapable of leaking you. It receives a claim ID, a proxy reply address, your phone's last 4, your state, and your DNC status. Not your name, not your email, not your address. A second layer then greps the generated letter against every representation of your real PII and hard-fails if any appears. That layer should never fire — which is precisely why it exists. The scammer gets a letter demanding money from someone they cannot identify. Status The full spine runs on live model calls: 66/66 verification checks pass, and the whole path — screen → capture → analyze → draft → approve → delivered — has been clicked end to end. Honest scope: the data broker registry is fictional placeholder data, and letters land in an in-app mock inbox rather than real mail. Both are deliberate — naming a real company as holding your data is a factual claim, and wiring a mail transport is a separate decision.
ShieldGuardClaude, Next, React, TypeScript, NodeJS, Cryptogithub.com - #15formThe decisions above your job
Civiq
City of Ventura publishes agendas for 21 boards and commissions as PDFs scattered across a portal. If you want to know whether anyone is voting on something near your house, you have to already know which of the 21 bodies to check, and then read the PDF packets. Most people don't, so they find out after the vote. Civiq lets you ask in plain language. Type "can they put a bar next to my house" and if it's an agenda item you get an agenda information back, each with a receipt: which body, the meeting date, the item number, the page range, and a link to the original PDF. You can also ask to be emailed when something matches. We ingested the whole live corpus, not a sample: 141 agendas across all 21 bodies, 707 items parsed, 0 failures. The design constraint we cared about most was that a model never invents a civic fact. Meeting dates are parsed from the URL. Item numbers and page ranges are read from the PDF's own text. The model's only job is rewriting staff prose into plain English. Four guardrails, enforced in code: 1. No model-generated dates, item numbers, or page ranges. 2. No scores, grades, or rankings about any agency. Nothing is ever called "late." Missing is phrased as "not located at [url] as of [timestamp]." 3. Nothing is ever sent to a government office. Email goes to residents only, and there is no send-to-city code path, not even a stub. 4. Every result links to the source PDF. Limitations we're stating rather than hiding: 6 of the 141 agendas are scans with no text layer. We store them with a link and flag them instead of running OCR and guessing. 22 have no parsable items, mostly cancellation notices, which we keep as published because a cancelled meeting is itself something a resident may be looking for. Ventura only. Agendas only, not minutes. Packet text is treated as untrusted data, never as instructions.
Team CiviqNext.js 16 (App Router) on Vercel · Neon Postgres with pgvector + tsvector · Claude Haiku 4.5 (rewrite) + Claude Sonnet 5 (chat) via Vercel AI SDK · Gemini embeddings, 768-d · unpdf for PDF text · Resend for emailgithub.com - #16formWildcard
Capital Compass
We built a tool for small business owners and entrepreneurs to find capital for their small business in a way that is approachable and easy to understand. This tool connects financial resources with those who may otherwise be unable to find or understand how to access them.
Capital Compassgithub.com - #17formWildcard
AssemblyLine
We built an interactive 3D visualizer that transforms traditional furniture assembly instructions into clear, step-by-step visual guides. Traditional instructions are often text-heavy, complex, and difficult to follow, making assembly frustrating for users. Our tool simplifies the process by visually showing how each component fits together, with interactive 3D models, animated tools, and guided build steps.
Wizard CookingClaude API, Javascript, React, HTML, CSS, Viteclaude-hackathon-mu.vercel.app