Today a local question gets answered three ways: one person pays for research, a journalist picks it up (rarely), or it dies in a neighbourhood WhatsApp group full of guesses. Patreon and GoFundMe fund people, not answers. Ask Together pools small amounts from the people who share the question and buys an AI investigation that every funder receives, with sources and confidence levels.
The data it runs on
Three sources. Web search through an LLM search tool (news, official notices, utility and municipal sites). Public records: open data portals where they exist, plus drafted right-to-information requests (India’s RTI Act, Ghana’s Right to Information Act 2019) that a funder can file with one tap. Local uploads: photos, notices and screenshots from residents, read by a vision model. Limits: many local bodies do not publish notices online, records requests take weeks, and AI cannot test water or measure a road. Every report therefore lists what is unknown. Fallback: when sources run dry, the report becomes a ready-to-send letter to the right office, signed by every funder.
1. The mission
Give every street a research desk. When something changes in a neighbourhood, the people affected should get a clear, sourced answer within an hour, for the price of a cup of tea each.
2. Why people can’t resist it
Curiosity with company. Everyone on the lane is already asking. Chipping in a little to finally know is irresistible.
Visible momentum. A progress bar filling with neighbours’ faces makes the 0.50 feel like joining something.
Collective power. The report ends with who to contact, and 30 signed complaints get a reply where one does not.
3. How it works for real
Login: email or phone; a passkey smart wallet is created behind the scenes (Coinbase Smart Wallet or Privy). Gas is sponsored by a paymaster on Base.
Pool: each question gets a small escrow contract holding USDC on Base. If the goal is not met in 7 days, everyone is refunded automatically.
Dedupe: a fast cheap model (Gemini Flash or DeepSeek via OpenRouter) merges duplicate questions within the same area, so neighbours fund one investigation.
Investigate: a frontier LLM with web search (Claude or GPT with a search tool, or Perplexity Sonar) plans queries, reads notices and news, and logs every source. A vision model (Gemini 2.5 Flash) reads resident photos. x402 pays per request for paywalled news or data APIs.
Verify: a second model checks that every claim cites a source and assigns confidence. Claims without sources are moved to "What is unknown".
Deliver: the report is unlocked for every funder’s wallet; follow-up questions draw from a small pool buffer.
4. Cost per use
About $2 per investigation: search and reading around $1.20, photo analysis $0.10, drafting and checking $0.60. Follow-ups about $0.04 each. Gas on Base is fractions of a cent and sponsored.
Free: read summaries of past investigations, ask questions, join a pool. Paid: funders pay what they chip in; a 15% platform fee is included in each goal. Deep tier (15 USDC) adds a local fact-checker who reviews the report. Councils, housing societies and local media can subscribe for alerts.
5. Growth loop
A question is born in a building WhatsApp group. The share link shows faces and a progress bar. The report is only for funders, so the summary card travels further with "Chip in to read the full report". Every report ends with a follow-up question, which seeds the next pool. Each Monday the Open questions board ranks the week's top-voted pledges, sends alerts to followers when a finding lands, and adds every answer to a searchable archive, so neighbours come back weekly to vote, pledge and read.
6. Risks and guardrails
Defamation: no naming private individuals; companies and officials only with sources; claims without sources are flagged as unknown.
Health and safety: water, food and building safety questions carry a standing note to contact the authority and never treat the report as a test.
Misinformation: confidence levels on every finding, sources linked, and a correction flow that notifies every funder.
Abuse: questions about a person are rejected; photo uploads have faces and number plates blurred.
7. MVP build order
Question board per area, email login, smart wallet, USDC escrow with goal and auto refund on Base Sepolia.
Investigation agent: web search plus source log, report template with confidence and unknowns.
Photo uploads read by a vision model.
Report unlock for funders, follow-up questions, share card.