Leadpoet
SN71AI sales agents that find and qualify your next customers automatically
Leadpoet runs a two-track sales intelligence subnet: one track where miners submit AI models that surface in-market companies for a buyer's ICP, and a second where miners deliver fully enriched leads on real, paid client requests through a commit-reveal flow. Validators score every lead through email verification, LinkedIn checks, company verification, and intent-evidence scraping before any reward lands.
// Sales pipeline, on-chain.
Leadpoet is a Bittensor subnet that turns the top of the sales funnel into a competitive marketplace. Miners compete to surface high-fit companies and to deliver verified contacts that match a buyer's ideal customer profile. Validators run the same independent scoring pipeline on every submission, so a lead only earns when it actually passes verification.
The simple version: It's like outsourcing your outbound sales research to a competitive network instead of a single agency, with built-in checks so you only pay for leads that are real, in-market, and on-target.
Centralized equivalent: Think ZoomInfo, Apollo, or Clay, but rewards flow to whoever produces the best-verified leads on each request rather than to a single data vendor.
How it works:
- Miners run in two tracks. In Model Competition, they submit a Python model that takes an ICP description and returns up to 5 matching companies with intent evidence. In Fulfillment, they answer real client requests with fully enriched leads (contact, role, verified email, intent signal) using a commit-reveal flow so other miners cannot copy their work.
- Validators run the same multi-tier scoring pipeline on both tracks: ICP fit, then data accuracy (email verification via TrueList and ZeroBounce, LinkedIn checks via ScrapingDog, company verification, reputation score), then intent scoring (each evidence URL is fetched, the snippet is matched, and an LLM verifies relevance).
- The problem it solves: Outbound sales data goes stale fast and most lead lists are stuffed with people who already left or never matched. Verification is expensive, inconsistent, and usually done by the buyer after the fact.
- The opportunity: Sales intelligence is a real budget line at almost every B2B company. The pitch is a network that pays only for leads that pass the same checks a buyer would run themselves.
- The Bittensor advantage: Permissionless miner entry plus continuous, identical validator scoring on every lead. A miner who games the signals (fabricated intent text, stripped LLM disclaimers, fake dates) gets banned and blacklisted by design, not by trust.
- Traction signals: The main repo was last pushed today, with 27 listed contributors and commits in the past 24 hours covering both the Model Competition v2 ICP set and the Fulfillment two-stage TrueList verification flow. The subnet has an active client-facing product surface at leadpoet.com.
Category: Data Scraping and Archival | Centralized Competitor: ZoomInfo, Apollo, Clay, Crunchbase
Lead generation is one of the few places in the AI stack where the data quality problem is more painful than the model quality problem. Buyers will pay handsomely for a verified inbox match with a real intent signal attached, and walk away from a list of 10,000 names where 30% bounce. Leadpoet's bet is that the right way to produce verified leads is a permissionless competition where validators run the same checks the buyer would run, and the network only rewards leads that pass.
Mechanism:
Model Competition runs the company-discovery side. A miner packages a Python model under 200KB with a find_leads(icp) function, submits it to the gateway with a $10 TAO payment, and validators evaluate it against a set of 20 fresh ICPs (one per industry, refreshed daily). The model must return up to 5 distinct real companies per ICP with strict schema (no person-level fields allowed in this track) plus structured intent signals where the source must match the URL domain. The highest-scoring model becomes the "champion" and holds the crown, earning rewards until a new model beats it by the published threshold.
Fulfillment runs the contact-delivery side. Clients post fulfillment requests with a structured ICP (industry, target roles, seniority, geography, intent signals, required attributes, number of leads). Miners sources contacts off-chain, then submit hashed lead data within a one-epoch commit window (about 72 minutes) followed by a 15-minute reveal window where the actual lead data must hash-match the commit. Validators then score each revealed lead through four tiers: Tier 1 ICP fit, Tier 2 email and identity verification, Tier 2c required-attribute checks via Sonar web search, and Tier 3 intent scoring where every evidence URL is fetched and the snippet is verified verbatim against the page. Winning leads earn 0.05% of emission per epoch for 100 epochs, which works out to roughly five days of payout per winning lead.
The anti-gaming rules are unusually explicit. The README lists six "instant ban" patterns including stripping negative LLM assessments, fabricating dates, injecting fake intent signals, fabricating evidence text, bypassing LLM verification, and defaulting verification to pass. Models also run in a sandbox with a hardcoded library allowlist and runtime monitoring for blocked patterns like eval, exec, os.system, and access to .bittensor or .ssh paths.
On-chain, the picture is mixed and worth reading carefully. Alpha price is 0.00606 TAO with a market cap near 28,829 TAO. The pool holds 5,995 TAO of root liquidity, and root_prop is 0.16, which means most of the depth is organic alpha rather than protocol subsidy. Price is up 13.5% over 7 days and down 9.7% over 30 days, with 7-day net staking flow positive at about 372 TAO. Under Taoflow, sustained positive net flows are what drive a subnet's share of network emissions, but the current emission share is 0% and miner burn sits at about 82%, which means the EMA has not yet caught up to the recent inflow. If positive flows continue, the share follows on the EMA's lag.
TAO.app's about page describes Leadpoet primarily as a single-pipeline lead-sourcing subnet with batch validation. The actual main repository describes the current two-track architecture (Model Competition plus Fulfillment with commit-reveal). Treat the repo as authoritative for mechanism details.
- Market: Emission share is currently 0% under Taoflow, with about 82% miner burn. Net 7-day flows are positive at roughly 372 TAO, which is the right direction for emission share to recover, but until the flow EMA turns the subnet receives no TAO injection from the network.
- Execution: The product surface is complex, two evaluation tracks, four scoring tiers, sandboxed model execution, commit-reveal, anti-gaming detection. Each layer is a real code path that has to stay reliable. Recent commits show active work on exactly these surfaces, which is the right pattern but also signals the system is still maturing.
- Competition: Sales intelligence is a crowded centralized market with well-funded incumbents. Leadpoet's edge is verifiable scoring and pay-for-verified-leads economics, not raw data scale. Whether buyers prefer that tradeoff at scale is still an open question.
- API dependency: The validator stack leans on TrueList, ScrapingDog, OpenRouter, and several other paid APIs for email and identity verification. Outages or pricing changes upstream propagate directly into validator cost and scoring availability.
- Concentration: Top-100 Gini is 0.535 and HHI is 0.020, with the top 10 controlling about 35% of supply. Stake leans on a small set of large positions, which is normal at this size but worth tracking if flows turn.
Into the next one.