• 

  Dashboard Image

LeadFlow

Automated lead generation and outreach for business owners

Inspiration

Running a small business means wearing every hat. Finding new clients often turns into hours of repetitive work: searching for prospects, digging through websites for contact info, and sending outreach that still feels generic.

We built LeadFlow so a business owner can run the entire pipeline automatically: find potential clients, research them, reach out with context, and track results in one dashboard.

What it does

LeadFlow is an AI powered client acquisition dashboard that automates outbound outreach from start to finish:

1. Find prospects (Search)

   • Type a query like "restaurants in LA" or "local gyms in Manhattan"
   • LeadFlow uses Exa neural search with Gemini to return real targets with their actual websites
   • Gemini expands broad queries into specific subcategories to improve lead quality

2. Enrich leads (Research)

   • One click pulls useful context from the prospect's website with firecrawl:
     • decision maker names
     • emails and phone numbers
     • what the prospect does and why they might be a fit
   • Enrich scrapes the website for structured info

3. Computer Use to uncover alternative contact paths (Contact)

   • If a site does not clearly list an email or phone number, LeadFlow uses Cloud Based Computer Use with Browser-base to navigate like a human and contact the business:
     • clicks through pages like About, Team, Contact, Support, and Careers
     • opens menus and footer links
     • finds contact forms, booking links, chat widgets, and social profiles
     • completes captchas

4. Reach out automatically (Calls)

   • LeadFlow can place real phone calls through Twilio
   • ElevenLabs powers natural sounding voice
   • Deepgram nova 2 provides low latency speech to text
   • Gemini drives the conversation to qualify interest and push toward booking a follow up

5. Track and classify outcomes (Dashboard)

   • Live transcript streaming during calls so you can watch in real time
   • After each call, LeadFlow classifies outcomes:
     • Interested
     • Not interested
     • Unreachable
   • Filter by status, enrich in bulk, and monitor everything through a live activity log

Challenges we ran into

• Directory page problem

  • Traditional search often returns Yelp or aggregator pages instead of the prospect's real website
  • We switched to Exa, which consistently returns the prospect's actual domain

• Call latency

  • Early versions had noticeable dead air between the user speaking and the AI responding
  • We reduced latency by caching TTS audio, using faster STT, switching to a lighter real time model, and enabling streaming optimizations

• Keeping prospect to URL mapping correct

  • When bundling results for extraction, the association between each prospect and their URL could get lost
  • We restructured the pipeline so URLs stay attached end to end

• Local vs cloud based computer use

  • When experimenting with computer use methods, we were originally running it locally using stagehand
  • We switched later on to browser-base which was much faster and ran in the cloud

What we learned

• Speech to text model choice has a huge impact on responsiveness. Deepgram nova 2 was noticeably faster for real time calls
• In voice AI, small optimizations compound: audio format, model size, caching, and streaming all matter
• Neural search like Exa is the difference between getting real websites vs directory junk
• Computer use enables workflows APIs cannot: navigating messy sites and uncovering contact paths

Built With

• ai
• browser-base
• deepgram
• elevenlabs
• exa
• firecrawl
• google-gemini
• next.js
• tailwind-css
• twilio
• typescript
• vercel