Source-aware acquisition
Twelve connectors share one lead contract while each source keeps its own fetch strategy, rate limit, retry behavior, and hydration rules.
Back to Work
Lead-Prospecting Platform2026Solo build
Pulse
A personal acquisition workspace for reviewing leads from twelve different sources in one queue. It combines source-specific collection, email verification, saved searches, and outreach sequences around a keyword library tuned to the work I actually pursue.
The inbox after a fresh scan — leads from HN hiring, YC W26 companies, ProductHunt launches, and IndieHackers founders consolidated in one queue with quick filters and matched-skill highlighting.
12
live data sources
1,832
curated keywords
42
industry/role bundles
4-stage
email verifier
The Problem
Lead-gen SaaS isn't built for freelance devs
I needed one review queue for remote contracts, product roles, recently funded companies, and local businesses. The existing tools I tried were built around different acquisition workflows:
- Apollo, Hunter, Lemlist — optimised for SDR teams working from established company lists. My search starts earlier, with roles, funding events, product launches, and founder activity.
- LinkedIn / Wellfound — auth-walled, anti-scraping, and most listings filter out international candidates without surfacing the filter upfront.
- HN "Who is hiring?" threads — gold-standard signal but lives in monthly comment threads with hundreds of posts. No filtering, no dedup, no structured data.
- Local SMBs (cold outreach) — phone numbers and websites scattered across Google Maps, Yelp, Yellow Pages, Manta, BBB. Nothing consolidates them with verified contact emails.
So I built the tool I wished existed: a personal lead engine that pulls from twelve sources in parallel, filters by a keyword library tuned for fullstack work (including French/Tunisian variants for local job boards), verifies emails before they land in the inbox, and surfaces everything in a keyboard-first triage UI.
Lead pipeline
Signal enters messy. Opportunity leaves structured.
Pulse treats prospecting as a pipeline, not a list of scraped links. Each stage removes noise or adds context until a lead is useful enough to review, contact, and track.
Contact confidence
Syntax, MX, SMTP, and catch-all checks filter weak addresses before outreach and prefer contacts that match the company domain.
Opportunity model
Keywords, bundles, source context, and saved-search criteria turn raw posts into comparable opportunities in one inbox.
Outreach workflow
Reviewed leads move into sequenced contact and reply tracking without losing the source evidence that made them relevant.
Saved Searches
Run scans on your schedule
Each saved search picks its sources, keywords, locations, and frequency. Hourly, daily, weekly, or on-demand. The background scheduler dispatches scans, persists results to Supabase, and updates the inbox. Saved searches auto-pause if a source rate-limits.
Editor
Direct vs Cold mode
Two distinct flows in one editor. Direct mode (jobs you apply to) takes keyword chips. Cold mode (businesses to sell to) takes location autocomplete + niche bundles. Switching modes resets sources to sensible defaults so a Cold scan never accidentally runs against LinkedIn.
Bundle Picker
42 quick-add bundles
Each chip toggles a whole bundle of keywords on or off in a single click. State is computed: a chip is "active" when every keyword from that bundle is selected, "partial" when only some are. Tap an active or partial chip to remove only that bundle's keywords; manually-added ones stay.
🚀 Founding eng🤝 Contract🤖 AI/ML/LLM💸 Fintech⚡ Modern Jamstack🇹🇳 French/Tunisian
Power UX
Cmd+K palette · J/K nav · A/R triage
Built for the kind of triage where you process 50 leads in 10 minutes. Cmd+K opens a fuzzy palette across every page and opportunity. J/K navigates the inbox, A applies/contacts, R rejects, V marks viewing, O / Enter opens the original posting in a new tab. Every action is reachable from the home row.
Engineering Challenges
The hard parts I had to solve
Scraping 12 sites with 12 different defenses
HN runs on plain HTML. YC ships everything via a public Algolia search index. Google Maps requires Playwright with stable data-item-id attribute selectors that survive UI redesigns. ProductHunt and IndieHackers are React SPAs behind Cloudflare and need a Camoufox-backed stealth fetcher with wait_selector to pause for client-side hydration. I built a singleScraperEngine that picks the right fetcher tier per source, applies per-source rate limits, retries with exponential backoff, and exposes one uniform DirectLead contract upstream.
Stopping email contamination
Naive email extraction returns garbage: Stripe checkout buttons leave @stripe.com, Sentry trackers leave 32-char-hex pixel addresses, Intercom widgets leak support emails. The verifier I built runs four stages — syntax → DNS MX → SMTP handshake → catch-all probe — and ranks results by domain match against the company's own website. A 25-suffix blocklist filters known third-party CDNs and tracking subdomains, so what reaches the inbox is the actual founder/recruiter email, not a payment-widget side-effect.
Async pipeline that doesn't feel stuck
Sequential email extraction made the dock progress bar freeze at 20% for two minutes. I rewrote the pipeline around asyncio.Semaphore-bounded concurrency: 8 parallel email-extraction workers, 4 parallel detail-page fetches, with live progress callbacks emitting after every lead. A 20-lead scan went from ~200s to ~30s and the user always sees forward motion.
Internationalised keyword matching
Tunisian job boards (Tanit) post in French. Two-letter tech stems like ai, ts, qa false-positive on words like Anglais and Carts. I built a word-boundary regex matcher that treats ., +, and # as word characters (so next.js, c++, c# match cleanly), normalises French accents (NFKD + diacritic strip), and applies negative-match guards (commercial, vendeur, femme de ménage) to keep software-only results clean.
Inbox triage UX
150 leads/scan needs a UX you can fly through. Cmd+K opens a fuzzy palette across every page and opportunity. J/K navigates, A applies/contacts, R rejects, V marks viewing, O / Enter opens the original posting in a new tab and auto-flips the stage. Quick-filter chips show live counts before you click them, so you know exactly how many leads "Has email" or "Founding eng" will surface. Matched keywords are highlighted inline in each description so the relevant signal jumps off the page.
System architecture
Interactive product above. Long-running work below.
The architecture separates the fast review experience from source-specific collection and verification jobs, while one shared record model keeps the entire prospecting workflow coherent.
Next.js workspace
React, TypeScript, TanStack Query, and reusable interface primitives power fast triage, filters, editors, and command-driven navigation.
FastAPI workers
Python, asyncio, Pydantic, and source adapters handle scheduled acquisition and verification outside the interactive request path.
Supabase data layer
PostgreSQL, authentication, and row-level rules provide one durable home for saved searches, leads, contacts, and outreach state.
Tiered fetch system
Direct APIs, HTTP parsing, Scrapling, and browser automation are selected per source instead of forcing every target through one brittle scraper.
Need a system like this for your team?
I build production scrapers, workflow automation, and full-stack internal tools end-to-end. Available for senior remote contracts.