Manual prospecting found only a small number of companies each day, with the reasoning scattered across browser tabs and notes. Fully automatic outreach would create a different problem: nobody could see why a company qualified, and an inaccurate or poorly timed message could still reach a real person.
05 / Internal product · AI workflows · B2B
Leads Discovery Engine
Built for Shree Export
An internal research engine that works out who buys an exporter's components, finds those companies from public and trade data, maps their suppliers and decision-makers, verifies how to reach them, scores fit, and drafts an opening email for a person to review.
Let AI gather evidence, help score fit, and prepare a draft. Keep the decision to contact a company with a person who can review the evidence and change or reject the message.
The engine has built a database of 30,000+ companies, 90,000+ contacts, and 50,000+ emails, producing 14,000+ qualified leads. It is backed by a CRM with a 7-stage pipeline, tasks, sequences, deals, and territories, and by export market reports for five target countries.
The workflow is built, but lead volume, research time saved, and email deliverability still need to be measured during regular use. Private company, contact, and lead data are not shown.
01
User and problem
Who needed this product, and why.
A small marketing team that researches and qualifies prospects, with sales and management able to view progress, assignments, and lead quality without changing the working data.
The original request was to find more export leads. Instead of searching for competitors, I mapped Shree Export's products to 17 buyer profiles, such as transformer makers who need brass bushings, and built a pipeline around them. Four regional workers search the web, crawl company sites, and pull trade data from UN Comtrade, UK HMRC, and US customs shipment records to see who imports what and from whom. Claude researches each company, identifies which roles make buying decisions, and suggests an approach angle. The platform then finds public LinkedIn profiles, guesses likely email formats, verifies them against the mail server, scores the lead, and drafts a personalised email that a person reviews before anyone makes contact.
02
Important decisions
Decision 01
Search for buyers, not competitors
- Why it mattered
- Searching for brass and copper parts mostly returned other exporters and B2B directories, which are competitors, not customers.
- What I did
- Map what Shree Export makes to who consumes it. 17 buyer profiles link each product line to the industries that buy it, and those profiles generate the search queries and the trade-data lookups by HS code.
- What it required
- Building the buyer map took product and market research up front, and every new product line needs a new profile.
- The result
- Searches target real buyers such as switchgear, transformer, and plumbing manufacturers across countries, and trade records show which of them already import similar parts.
Decision 02
Find the person who decides, and check the way to reach them
- Why it mattered
- A company record with a generic info@ inbox rarely reaches anyone who can place an order.
- What I did
- Have Claude identify the buying roles for each company type, extract named contacts from the company site, find their public LinkedIn profiles through web search with strict name and company matching, generate likely email formats, and verify each one against the mail server with catch-all detection.
- What it required
- Verification is slower than guessing, and catch-all domains can only be marked as uncertain rather than confirmed.
- The result
- Each lead shows who to contact, why they matter, and how confident the platform is that the email will arrive.
Decision 03
Give every step a visible state
- Why it mattered
- One automated score would hide whether a company was only discovered, researched with evidence, judged as a fit, reviewed by a person, or contacted.
- What I did
- Keep companies, contacts, source evidence, qualification, assignment, status, and outreach approval as separate records and steps.
- What it required
- The team has to make a few explicit decisions instead of pressing one automation button, but each decision can be checked and corrected.
- The result
- Anyone with access can see where a lead is, why it qualified, who owns the next action, and which person approved the outreach.
Decision 04
Build the team workspace before heavier AI
- Why it mattered
- The long-term concept included continuous research, monitoring, reporting, and campaigns, which was too much to validate in one release.
- What I did
- Start with login, role-based access, companies, leads, job status, filtering, assignment, bulk actions, and export. Add deeper AI research only after the team can manage the resulting work.
- What it required
- The first release automates less, but gives the team one reliable place to organize and review prospecting work sooner.
- The result
- The working dashboard supports the daily process now, while the roadmap keeps later research-agent capabilities separate from features already built.
Decision 05
Draft with AI, contact by a person
- Why it mattered
- AI-generated personalization can be inaccurate, intrusive, or poorly timed when it is sent without review.
- What I did
- Save each AI draft on the lead record for a person to read and edit. The platform never sends email by itself.
- What it required
- The workflow sends fewer messages than a fully autonomous system, and every approved draft requires human attention.
- The result
- A wrong or badly timed message cannot reach a real company unless someone has read it and chosen to approve it.
03
What I built
What the product includes and how it works.
- 17 buyer profiles that turn product lines into targeted searches and HS-code trade lookups
- Four regional workers (US, EU, Asia, MENA) plus an autonomous agent that tunes queries based on results
- Trade intelligence from UN Comtrade, UK HMRC, and US customs shipment records to see who imports what and from which suppliers
- Claude company research covering fit, buying roles, approach angle, and competition risk
- Contact discovery: site extraction, public LinkedIn profile matching, email format generation, and mail-server verification with catch-all detection
- Scoring across company fit, contact quality, and email confidence
- A role-based internal dashboard for marketing, sales, and management
- Company and lead records with status, assignment, filtering, bulk actions, and CSV export
- Structured source evidence and qualification fields that show why a lead was accepted or rejected
- Visible background-job status for longer research and enrichment tasks
- AI-drafted outreach saved on each lead for a person to review; nothing is sent automatically
- A CRM layer with a 7-stage pipeline, tasks, follow-up rules, email sequences, deals, and territories
- A PRD and phased roadmap that keep future AI-agent work separate from the working MVP
How it works. The web dashboard stores companies, contacts, public source evidence, qualification decisions, assignments, and outreach drafts as related records with access controlled by role. Longer research and enrichment tasks run in the background with visible status. AI-generated drafts are saved on the lead record, not sent, so contact only happens when a person decides to make it.
04
What exists today
30,000+ companies, 90,000+ contacts, 50,000+ emails, and 14,000+ leads in the database
Totals across all discovery runs to date. Company names, contacts, and emails are kept private.The manual process found about 5 to 10 leads per day
This recalled working estimate was the starting point for the product, not an instrumented result.Goal: build a list of 500 qualified leads per month
This is a product goal, not a result achieved yet.Goal: reduce research time and improve email deliverability
Time saved and the share of messages reaching an inbox still need to be measured during regular use.Private names, contact details, and lead data are not shown
The case explains the product without exposing business or personal data.05
Reflection
What I learned, and what I would test next.
- An AI workflow is easier to inspect and improve when each step records a visible decision instead of hiding activity inside an agent.
- Product goals must stay clearly separate from results that have actually been measured.
- When AI can affect an external person, human review can be a permanent product safeguard rather than a temporary limitation.
Let’s discuss the product problem.