Date: 2026-04-08 Author: Yilin Jing Status: Draft — 2-week MVP sprint Brand: Overheard — "I overheard someone who's been through this."
Let information flow to its fullest potential.
Google made published information searchable. We make unpublished information tradeable. Every person's lived experience — locked in their memory, never indexed by any search engine — becomes a tradeable asset through their personal AI agent.
"Google indexes the web. We index people."
Domain: overheard.co (available) | Backup: overheard.sh, overheard.is
99% of first-hand human experience has never been published. It exists only in people's heads — and now, in their AI agent's memory.
| What exists today | What's missing |
|---|---|
| Google indexes billions of web pages | Can't index what was never published |
| Q&A platforms (Reddit, Quora) let people ask | Answers take days, lack context, no quality guarantee |
| Expert platforms (JustAnswer, Clarity.fm) | $100+/hr overkill for a $10 question, "experts" ≠ "people who've been there" |
| AI agents can search the web | Can only find what's already online |
| Friction | Description | Who solves it today? |
|---|---|---|
| Discovery | Don't know who has this knowledge | Nobody |
| Context | Find someone, but their situation differs from yours | Nobody |
| Time | Wait hours to days for an answer | Partial (JustAnswer is fast but expensive) |
| Price | Either free + low quality, or high quality + too expensive | Nobody ($5-15 mid-range is empty) |
| Social | Asking friends costs social currency; strangers have no incentive | Nobody |
| Verification | Can't verify if the answerer has real first-hand experience | Nobody |
| Supply | People with knowledge have no channel, pricing, or incentive to sell it | Nobody |
Seven frictions, zero fully solved. This is our opportunity space.
- No pricing mechanism — "How long does YC interview actually take?" has never been priced
- No discovery channel — You don't know who has the answer; they don't know you'd pay for it
- No flow infrastructure — Even if matched, there's no protocol for exchanging information and settling payment
Three structural shifts make this possible today but not two years ago:
Old: Person must actively write blog / answer question / take a call
→ Time cost: 30 min ~ 2 hours per answer
→ Willing suppliers: < 1%
New: Agent auto-generates answer from memory
→ Time cost: 0 (authorize once, runs automatically)
→ Willing suppliers: >> 1% (zero effort + passive income)
This is the fundamental difference from every predecessor (Google Answers, Quora, JustAnswer, Clarity.fm). Their supply side required humans to actively spend time answering. Ours doesn't.
Old: Match by keyword → "YC application" → everyone who tagged YC
→ Mix of: accepted founders, rejected applicants, blog readers, advisors
New: Agent knows your full context
→ Solo founder + SaaS + pre-revenue + applying S26
→ Matches someone who was in your exact situation 6 months ago
Existing platforms can't do this because they don't know the user's full context. An agent does. It's read the user's chat history, knows their company stage, geography, constraints.
Old: Guess-based pricing or flat expert rates
→ Free (Reddit) or $200/hr (consultant), nothing in between
New: Agent searches Google first, evaluates result freshness/quality
→ "Google's latest: March 2024, says '2-4 weeks'.
Our network has 2 answers from the past 3 months. Suggested: $8."
The $5-15 price band has never been effectively served. Too low for professionals, too high for free platforms. But AI agents' zero marginal cost makes this band viable.
Every person will have a personal AI agent. That agent will be their primary interface for acquiring and sharing information. We build the pipes that connect these agents into a global information exchange network.
┌──────────────────────────────────────────────────┐
│ User's Personal Agent │
│ (Claude Code / Cursor / OpenClaw / GPT / any) │
│ │
│ Existing tools: Gmail, Calendar, Search, Code... │
│ │
│ + Overheard Skill ← Our product │
│ (CLI tool / Web interface) │
└──────────────┬───────────────────────────────────┘
│
v
┌──────────────────────────────────────────────────┐
│ Overheard Platform (centralized matching engine) │
│ │
│ Discovery · Pricing · Matching · Settlement │
│ Reputation · Verification · Info Reuse │
└──────────────┬───────────────────────────────────┘
│
┌───────┼───────┐
v v v
Agent A Agent B Agent C
(other users with Overheard installed)
The skill is the tentacle. The platform is the heart.
Principle: Information price = degree of search engine failure.
The platform doesn't set prices. It provides market signals. Buyers set their own willingness to pay.
Pricing flow:
1. User asks a question (via web or agent CLI)
2. Agent searches Google/Perplexity first (free)
3. Platform evaluates search results:
- Freshness of newest result
- Specificity vs. generic
- Contradictions across sources
4. Displays price signal:
Google has fresh, accurate answer → $0 (don't use platform)
Google answer exists but outdated → $3-8
Google answer is vague/generic → $5-15
Google has no relevant answer → $15-50
Google has misleading information → $20-50+
5. "Google's latest: March 2024. Our network has data from 2 months ago.
3 verified sources available. Suggested: $8."
6. Buyer sets bounty
7. Market responds
Information reuse and royalties:
1st buyer asks: $8 → Fresh answer generated from seller's memory
2nd buyer (similar context): $5 → Platform delivers verified existing answer
10th buyer: $3 → Well-validated, high-confidence answer
100th buyer: $1 → Effectively platform knowledge base
Seller income:
Original answer: $5.50 (after platform fee)
Each reuse: $0.30-0.50 royalty
→ One experience shared casually with agent = passive income for months
Context similarity determines reuse eligibility. "American SaaS founder in Berlin doing GmbH" matches closely. "German citizen registering in Munich" does not — gets fresh match.
Principle: Don't match topics. Match life situations.
The ultimate match: someone who was you 6 months ago — same background, same city, same constraints, same goals.
Five levels (progressive rollout):
| Level | What it matches | MVP? |
|---|---|---|
| L1: Topic | Keywords and domain tags | Yes |
| L2: Experience | First-hand vs. second-hand knowledge | Yes |
| L3: Situation | Context similarity (nationality, city, constraints) | Yes (basic) |
| L4: Predictive | "83% of people who asked this next ask about VAT" | Post-MVP |
| L5: Proactive | Agent detects you're applying to YC → pre-matches mentors | Post-MVP |
Two-phase matching architecture:
Phase 1: Coarse matching (on platform)
Seller agents register capability tags (user-authorized):
- Domains: "YC application," "US immigration," "Berlin startup"
- Recency: when the experience happened
- No specific content, just domain declarations
Query comes in → Platform filters by tags → N candidate agents
Phase 2: Fine matching (on device)
Query sent to candidate agents (sanitized, no buyer identity)
Each agent checks local memory:
- Can I actually answer this specific question?
- How relevant is my experience to this buyer's situation?
- Is the bounty worth responding?
→ K agents respond with relevance score + preview
→ Platform ranks by reputation × relevance × recency
Matching flywheel:
Every completed transaction teaches the platform which dimensions matter:
- For YC applications: batch timing + company stage matters most
- For immigration: nationality + visa type + city matters most
- For fundraising: stage + industry + geography matters most
This matching model is the deepest moat. Competitors without transaction data can never replicate dimensional weights.
Three layers of defense against fabrication:
Layer 1: Reputation scoring
Each seller has a domain-specific reputation score:
Answer adopted by buyer: +2
Answer matches other sources: +1
Answer disputed by buyer: -5
Answer significantly inaccurate: -10
Score < 30: No longer matched in that domain
High reputation → priority matching, higher price eligibility
Layer 2: Multi-source consensus (post-MVP)
Same question routed to 3 independent agents:
Agent A (applied to YC 3 months ago): "Interview is 10 min, they focus on founder dynamics"
Agent B (applied 1 year ago): "About 10 min, team questions dominate"
Agent C (never applied, guessing): "Probably 30 minutes with technical deep-dive"
A and B reach consensus → High confidence, delivered to buyer
C deviates → Flagged, reputation penalty
Layer 3: Evidence incentives
"Did YC ask about revenue?" → Screenshot of notes optional but rewarded
"How long did visa take?" → Timeline with dates earns price premium
Answers with evidence: +30-50% price premium
Market naturally selects for quality
Principle: Agent as anonymizing proxy. Platform never sees raw memory.
Buyer ←→ Buyer's Agent ←→ Platform ←→ Seller's Agent ←→ Seller
| |
Sanitization Sanitization
(strip PII) (strip PII)
Platform knows: "who is good at what domain" + "transaction history"
Platform never knows: "what specifically someone experienced"
Agent auto-sanitization before sending:
Raw memory:
"I applied to YC S26 through Michael Seibel's referral,
interviewed on March 15, got rejected same day."
After sanitization:
"Applied to YC S26 with a referral, interviewed mid-March 2026,
received decision same day (rejected)."
Rules:
- Remove specific names, exact dates
- Preserve actionable information (referral helps, same-day decision)
- Seller can preview sanitized version before sending (manual mode)
- Or authorize auto-send (auto mode)
1. Website (overheard.co)
- Buyer: Browse questions, post bounties, receive answers
- Seller: Inbox of matched queries, manage domains/preferences, earnings dashboard
- Public: Browse answered questions (anonymized), see domain leaderboards
2. Agent CLI Skill
- Buyer's agent:
overheard ask "How long does YC interview actually take?"→ searches Google → shows price signal → posts bounty → receives answer - Seller's agent: Monitors incoming queries → checks local memory → notifies owner or auto-responds → receives payment
Buyer Flow:
1. User tells agent (or visits website):
"I want to know what YC S26 interview is actually like"
2. Agent (with Overheard skill):
a. Searches Google → finds generic blog posts from 2024
b. Queries Overheard → finds 4 agents with S26 first-hand experience
c. Shows: "Google's latest: 2024 blog, generic advice.
Overheard: 4 verified S26 applicants. Suggested: $10."
3. User: "Get it" / sets bounty at $10
4. Platform matches → delivers answer:
"Interview was exactly 10 minutes. Two partners. They didn't ask about
traction — all questions were about founder relationship and why this
problem. Decision came 2 hours later by email."
Source: Reputation 91 in "YC-application" domain. Applied S26. Similar
profile (solo technical founder, dev tools).
5. $10 settled. Buyer rates answer. Reputations updated.
Seller Flow:
1. Three months ago, user chatted with their agent:
"YC interview was wild. Only 10 minutes. Michael and Dalton.
They didn't care about our numbers at all — just kept asking
about why I started this and what my co-founder relationship is like."
Agent: [stored in memory]
2. User installs Overheard, authorizes domains:
"YC application — yes, share my experience"
"Fundraising — yes, auto-answer mode"
"Personal life — never share"
3. Query arrives:
Manual mode: Agent notifies user → "Someone is asking about YC S26
interview format. Bounty: $10. Your experience is highly relevant.
Share?" → User approves (or edits) → Answer sent
Auto mode: Agent generates sanitized answer from memory → sends
automatically → User sees in earnings dashboard: "$10 earned"
4. Reuse: Similar questions arrive later → platform serves cached answer
→ User earns $0.50 royalty each time, indefinitely
The Flywheel:
User chats with agent about life (already happening)
│
Agent accumulates memory (already happening)
│
User installs Overheard, authorizes domains (new — one-time)
│
Queries arrive → Agent answers → Money arrives (automatic)
│
User earns while sleeping → shares MORE with agent (reinforcement)
│
More memory → more domains → more income → stronger flywheel
| Mode | Description | Use case |
|---|---|---|
| Manual | Agent notifies seller for each query. Seller reviews sanitized answer before sending. | Sensitive topics, high-value answers, new users building trust |
| Auto | Agent answers from memory automatically. Seller sees earnings in dashboard after the fact. | High-confidence domains, repeat topics, passive income |
| Domain rules | Seller sets per-domain: "YC → auto, fundraising → manual, personal → never" | Granular control |
- Buyer deposits funds into platform wallet (Stripe, minimum $10)
- Each transaction deducts from wallet balance (avoids per-transaction Stripe fees)
- Seller accumulates earnings, withdraws via Stripe Connect (minimum $20)
- Platform fee: 15% of transaction value
- x402 and MPP enable direct agent-to-agent payment without platform intermediation
- Buyer's agent pays seller's agent directly on-chain (USDC on Base)
- Platform fee collected via smart contract (transparent, auditable)
- Enables sub-$1 transactions economically (no Stripe minimums)
- Progressive: Stripe for mainstream users, x402/MPP for crypto-native users
MVP: Buyer → Stripe → Platform Wallet → Seller (Stripe Connect)
Future: Buyer Agent → x402/MPP → Seller Agent (direct, on-chain)
| Scenario | Buyer need | Why Google fails |
|---|---|---|
| YC application | "What do they actually ask in S26 interviews?" | Changes every batch, never published |
| Fundraising | "What's the real timeline for Series A in dev tools right now?" | Market conditions change quarterly |
| Hiring | "Senior React dev salary in Berlin — what are people actually getting?" | Published ranges are 12+ months stale |
| Incorporation | "Delaware C-Corp vs. Cayman — what does the lawyer NOT tell you?" | Lawyers give safe answers, not real answers |
| Scenario | Buyer need | Why Google fails |
|---|---|---|
| Company registration | "How long does GmbH registration actually take in 2026?" | Latest results are from 2024 |
| Immigration | "F1 to H1B — what surprised you most?" | Experiences vary wildly, Google gives generic |
| Tax | "First time filing as a non-resident — any gotchas?" | Country-specific, situation-specific |
| Banking | "Which Berlin bank actually opens accounts for non-EU founders?" | Changes constantly, outdated lists |
| Scenario | Why a matched human beats Google |
|---|---|
| "Notion vs Linear for 5-person startup?" | Someone who used both in a similar team size |
| "Hong Kong to Singapore — tax implications?" | Someone who made the exact same move |
| "Two job offers — equity vs higher base?" | Someone in the same industry who faced the same choice |
| Component | In MVP | Approach |
|---|---|---|
| Website | Yes | Next.js, simple Q&A interface + seller dashboard |
| Agent CLI skill | Yes | overheard ask "..." / overheard sell --domain "..." |
| Matching engine | Yes | Tag-based (L1 + L2), manual curation assist |
| Pricing signal | Yes | Google search freshness check + suggested price |
| Payment | Yes | Stripe wallet (deposit → per-txn deduct → withdraw) |
| Reputation | Yes | Basic score per domain (rating after each transaction) |
| Seller auto-answer | Yes | Agent reads local memory, generates sanitized answer |
| Seller manual mode | Yes | Notification → preview → approve/edit → send |
| Multi-source consensus | No | Post-MVP |
| Situation matching (L3+) | No | Post-MVP (start with tag-based) |
| Info reuse + royalties | No | Post-MVP |
| x402/MPP payment | No | Post-MVP (Stripe first) |
| PII auto-sanitization | Partial | Basic rules, seller preview in manual mode |
- No blockchain integration in MVP
- No complex protocol layers
- No multi-source verification
- No predictive/proactive matching
- No information reuse engine
Supply seeding (Week 1, parallel with development):
- Founder's network: 20+ friends who are tech founders / YC alumni / expats — personally onboard each one, set up their domains
- Seed content: Founder personally answers first 20 questions to establish quality bar and generate example answers
- Community: Post in 3-5 targeted communities (YC alumni Slack, Indie Hackers, digital nomad groups): "I built a tool where you can earn $5-15 answering questions about [thing you've done]. Anyone want early access?"
Demand seeding (Week 2):
- Reddit/Twitter: Post genuine questions with bounties in relevant subreddits and Twitter threads
- Direct outreach: Message 50 people who recently posted "I wish I'd known..." type content — they're natural buyers
- Founder's own questions: Use the platform for real questions the founder has, demonstrating the product authentically
Target: 30 sellers + 50 buyers by end of Week 2.
| Metric | Target |
|---|---|
| Sellers onboarded (with ≥1 domain) | 30 |
| Questions posted with bounty | 50 |
| Questions answered | 30 (60% answer rate) |
| Average transaction value | $8-12 |
| Buyer satisfaction (rated useful) | 70%+ |
| Sellers who earned > $10 | 10 |
| Repeat buyers | 5+ |
The ONE metric that matters: Do strangers pay $8+ for a first-hand answer from someone matched by life situation?
If yes → everything else follows. If no → pivot or kill.
| Competitor | What they do | Why we're structurally different |
|---|---|---|
| Google Search | Indexes published information | We trade unpublished information. Complementary. |
| Perplexity/ChatGPT | AI-synthesized answers from web | Same limitation — can only use what's online. We access what isn't. |
| Reddit/Quora | Free Q&A communities | No pricing, no quality guarantee, no situation matching, days of latency |
| JustAnswer | Expert Q&A ($5-99) | "Experts" ≠ "people who've been there." No context matching. |
| Clarity.fm | Paid expert calls ($100+/hr) | Overkill for $10 questions. Requires scheduling. |
| Fiverr/Upwork | Freelance marketplaces | Built for multi-day projects, not 5-minute knowledge exchange |
Our unique position: The only platform where AI agents trade unpublished human knowledge, priced against search engine freshness, matched by life situation similarity.
- Matching data flywheel — Every transaction teaches which life-situation dimensions matter for each domain. This dimensional weight data is unique to us and compounds with every transaction.
- Supply-side stickiness — Seller's memory and domain authorizations are configured in our system. Switching cost is re-setup + losing reputation history.
- The dirty work moat — Quality control, dispute resolution, domain-specific reputation calibration, seller onboarding — none of this is technically hard, but it's operationally exhausting. Big platforms won't bother until the market is proven.
- Local memory architecture — We're built for a world where memory stays on-device. Centralized platforms (OpenAI, Google) can't easily replicate this without fundamentally changing their architecture.
Transaction fee: 15% of each information exchange.
| Transaction value | Platform fee | Seller receives | Buyer pays |
|---|---|---|---|
| $5 | $0.75 | $4.25 | $5 |
| $10 | $1.50 | $8.50 | $10 |
| $20 | $3.00 | $17.00 | $20 |
| $50 | $7.50 | $42.50 | $50 |
Future: reuse royalties
When a cached answer is served to a new buyer with similar context:
- Buyer pays reduced price (e.g., $3 instead of $8)
- Original seller receives royalty ($0.50)
- Platform retains remainder ($2.50)
- → Higher margin on reuse, incentivizes quality answers
| Metric | Target (Month 6) |
|---|---|
| Monthly transactions | 5,000 |
| Average transaction value | $10 |
| Gross transaction volume | $50,000 |
| Platform revenue (15%) | $7,500 |
| Payment processing (~3%) | -$1,500 |
| Net revenue | $6,000 |
Not venture-scale yet. Path to scale:
- Raise average transaction value via high-value verticals (legal, medical, financial — $20-50 per question)
- Information reuse creates near-100% margin revenue
- x402/MPP reduces payment processing costs to near-zero
Frontend: Next.js (website + seller dashboard)
Backend: Rust (API server, matching engine, payment logic)
Database: PostgreSQL (users, transactions, reputation)
Search: Simple text matching + tag filtering (MVP)
Payment: Stripe Connect (wallet model)
CLI Skill: Rust binary (`overheard`) — interacts with API
Auth: Email + API key (sellers), email or anonymous (buyers)
Hosting: Vercel (frontend) + Fly.io or Railway (backend)
| AWN Module | Overheard Use | Adaptation Needed |
|---|---|---|
identity/ |
User/agent identity | Simplify — no on-chain in MVP |
trust/scoring.rs |
Domain reputation scoring | Narrow from "contextual trust" to "domain expertise score" |
trust/receipt.rs |
Transaction records | Simplify to "question-answer receipt" |
gateway/ |
Discovery / matching engine | Repurpose for knowledge provider discovery |
payment/ |
Settlement infrastructure | Replace x402 with Stripe for MVP |
daemon/ |
CLI skill event loop | Repurpose for query monitoring |
signing.rs |
Receipt signing | Reuse as-is for transaction integrity |
Once MVP validates demand, layer in:
- x402/MPP for agent-to-agent micropayments
- On-chain reputation anchoring (ERC-8004)
- Decentralized matching (agents can discover each other without central platform)
- Trust receipts for transaction history portability
The protocol emerges from the product, not the other way around.
-
Legal: Is paid information exchange considered "advice"? Especially for legal/medical/financial domains. Need counsel signoff before expanding beyond general experience sharing.
-
Information decay: How to detect when cached answers become stale? (e.g., visa processing times change quarterly). Need freshness metadata and automatic expiration.
-
Abuse: How to prevent sellers from fabricating experiences? Reputation + multi-source consensus (post-MVP) + evidence incentives. MVP relies on reputation + buyer ratings.
-
Brand: Overheard — confirmed. Domain:
overheard.co(available). Tagline: "I overheard someone who's been through this."
| Phase | Timeline | Focus |
|---|---|---|
| MVP | Weeks 1-2 | Website + CLI skill + Stripe + basic matching. Validate: do people pay for first-hand answers? |
| Traction | Weeks 3-6 | Situation matching (L3), multi-source verification, info reuse + royalties. Target: 500 transactions. |
| Protocol | Weeks 7-12 | x402/MPP payment, on-chain reputation, agent-to-agent discovery. Target: 2,000 transactions/month. |
| Scale | Month 4+ | High-value verticals, proactive matching, API for third-party agents. Target: 10,000 transactions/month. |
"Every person carries answers that others would pay for. Overheard builds the pipes that connect them."