Case Study Live at lead-detective.com ↗ Enterprise Sales AI $75K / yr

Lead Detective. One week of enterprise
sales research. Eight minutes.

An AI pipeline that pulls from 19 scrapers and data sources — LinkedIn, SEC EDGAR, Crunchbase, and more — to generate executive dossiers, account plans, and conversation guides automatically. Enterprise teams walk into high-stakes calls prepared.

8 min
Per dossier (vs 3-5 days)
19
Scrapers & data sources
$75K
Enterprise license/year
85-200x
Cost advantage vs human research

Enterprise sales teams walk into
$100K deals underprepared.

Quality pre-call research — understanding an executive's background, career arc, leadership style, company strategy, and conversation hooks — takes 3 to 5 days when done by a human analyst. Most teams skip it. Lead Detective eliminates the tradeoff.

The Research Gap
MetricHuman AnalystLead Detective
Time per dossier3 to 5 days~8 minutes
Cost per dossier~$1,500~$12.50 to $17.85
Research depthSurface-level web searchCareer trajectories, LinkedIn posts, SEC filings, rapport hooks, tech stack, funding history
🔍

Executives Are Researched Shallowly

A 10-minute Google search before a call does not surface career trajectories, published writing, board positions, or the topics an executive cares about publicly. It surfaces their LinkedIn headline.

Research Takes Time Teams Do Not Have

A team of 10 SEs spending 10 hours per week on research equals 400 hours per month — $720K per year in labor. That time either gets cut (bad calls) or gets spent (no selling time).

📋

No Tool Prepares The Conversation

ZoomInfo tells you who to call. 6sense tells you when to call. Nobody tells you what to say or how to open the conversation with the specific person in front of you. Lead Detective fills that gap.

Six automated steps.
Zero manual research.

Every investigation runs the same deterministic pipeline. User pastes a name. Eight minutes later they have a dossier a human analyst would need a week to produce.

Step 01 Company Setup Wizard

User pastes their company website. AI automatically researches the company, identifies ICP, target personas, pain points, differentiators, and value propositions. Configures lead scoring criteria and research prompts. No manual setup required — the wizard does it.

ICP definition Persona mapping Pain point extraction Value prop generation
Step 02 Lead Discovery

AI-powered prospecting finds companies matching the ICP. Three discovery modes: ICP Match (ideal profile companies), Competitor Customers (accounts using alternatives you should displace), and Buying Signals (recent funding, key hires, expansion signals). Each lead gets a fit score 0-100 with machine-generated reasoning.

ICP Match mode Competitor Customers Buying Signals Fit score 0-100
Step 03 Lead Scoring

Pass/fail gates plus weighted signifiers. Hard requirements filter out disqualified accounts first. Soft signals weight the remaining accounts. Outputs four tiers: Hot, Warm, Nurture, Disqualified. Every lead has a score with the reasoning surfaced in the UI so reps understand why a lead ranked where it did.

Hard requirements gate Weighted signifiers Hot / Warm / Nurture / Disqualified
Step 04 Deep Company Research

Multiple scrapers fire in parallel: LinkedIn company page, tech stack via BuiltWith, SEC EDGAR filings, Crunchbase funding rounds, Glassdoor sentiment, Google News, company website. Raw data is structured into XML blocks with data-tier labels. AI writes a company profile, account plan, and org structure from the structured data.

LinkedIn Company BuiltWith tech stack SEC EDGAR Crunchbase Glassdoor Google News
Step 05 Executive Research

LinkedIn profile scrape: title, career arc, skills, tenure. Recent LinkedIn posts for personalization hooks. Twitter/X activity. Published articles. AI synthesizes a person profile with conversation topics, rapport hooks, and explicit notes on what to avoid. The output reads like a briefing written by an experienced SDR who spent three days on the account.

LinkedIn Profile LinkedIn Posts Twitter/X Published articles Rapport hooks Conversation guide
Step 06 Hero Identification

Automated contact ranking by buyer likelihood. Title matching carries 60% weight. Seniority bonuses: C-suite +15%, VP +10%. Recency signal: new in role within two years adds +10%. Threshold of 70 or above designates a contact as Hero — the person most likely to be your economic buyer or champion at that account.

Title matching (60%) C-suite +15% VP +10% New in role +10% Hero threshold: 70+

Claude running as a subprocess.
Not a simple API call.

The most important architectural decision in Lead Detective: Claude AI is spawned as a child process with web_search and file_write tools — the same workflow as a human analyst. It reads data, searches the web for gaps, and writes documents. It is not prompted and responded to. It works.

Apify Actor Network

19 specialized scrapers and data sources running across the research pipeline. Each scraper is purpose-built for its source: LinkedIn Profile, LinkedIn Company, LinkedIn Posts, BuiltWith tech stack, Google News, Glassdoor, SEC EDGAR, Crunchbase, Twitter/X, and more. Data quality filtering and deduplication run before the AI sees anything.

🔗 LinkedIn Profile 🏢 LinkedIn Company 📝 LinkedIn Posts 🔧 BuiltWith 📰 Google News ★ Glassdoor 📄 SEC EDGAR 📊 Crunchbase 🐦 Twitter/X
Tiered Data Integrity

Three data tiers. Tier 1 is Apify-scraped primary data: LinkedIn, SEC, Crunchbase. Tier 2 is structured company data from websites. Tier 3 is AI web search fill-in. Tier 1 data is never overridden or contradicted by Tier 3. The hierarchy is programmatically enforced in the XML structuring layer before Claude sees the data.

Tier 1
Apify scraped (LinkedIn, SEC, Crunchbase) — authoritative, never overridden
Tier 2
Structured company data from web — fills gaps
Tier 3
AI web search — fills remaining gaps, never overrides Tier 1
Effect-TS Concurrency

Maximum 5 simultaneous research jobs running at any time. 10-minute timeout per job with graceful failure handling. Effect-TS manages the concurrency model so no job queue can starve or deadlock. SSE (Server-Sent Events) stream real-time progress updates to the frontend as each pipeline step completes.

Self-Configuring Setup

The setup wizard researches your company and builds your full ICP automatically. No consultants, no onboarding calls. Paste your website and the system configures lead scoring criteria, research prompts, and persona definitions without human input. One project can have multiple revenue channels with separate personas and value propositions.

Independent Verification Layer

A second, independent model fact-checks every material claim in the dossier with a fresh web search before it reaches the rep. Each claim is graded High, Medium, or Low confidence. Anything that cannot be confirmed is explicitly labeled unverified rather than presented as fact.

Frontend
Next.js 15/16React 19TypeScriptTailwind CSSshadcn/ui
Backend
Next.js API RoutesEffect-TSSSE streaming
Database
Supabase PostgreSQLLeadDet_ prefixed tables
AI & Scraping
Claude API (subprocess)web_search toolfile_write toolApify network (19 actors/sources)
Deployment
Render.com (backend)lead-detective.com (frontend)Max 5 concurrent jobs10-min timeout per job

$75K per year.
85 to 200x cost advantage.

$75K
Enterprise License
350 deep investigations per month included
~$15
Cost Per Dossier
$12.50-$17.85 (AI compute + scraping) vs $1,500 human equivalent
100x
Cost Reduction
Per dossier vs senior SDR time at $1,500/dossier
The ROI Math
Current cost (10 SEs)
$720K/yr
10 SEs x 10 hrs/week research x 52 weeks x $138/hr blended
Lead Detective license
$75K/yr
350 deep investigations per month, unlimited users
One additional deal
$200K/qtr
Closing one additional $200K deal per quarter = 2.5x payback on license alone
Competitive Positioning
ToolAnswersWhat It Misses
ZoomInfo / ApolloWHO to callDoes not tell you what to say
6sense / DemandbaseWHEN to callDoes not prepare the conversation
Lead DetectiveWHAT to say ← IPS fills this gapNothing. This is the missing layer.
Gong / OutreachHOW the call wentPost-game analysis, not pre-game preparation

Five engineering problems
that no off-the-shelf tool handles.

01
Claude as a subprocess — not an API call
Lead Detective does not call the Claude API and wait for a JSON response. Claude is spawned as a child process with web_search and file_write tools. It behaves like a human analyst: reads the structured data, identifies gaps, searches the web for missing pieces, writes draft documents, and iterates. The workflow is deterministic but the execution is agentic. This required building a subprocess orchestration layer with stdin/stdout handling, timeout enforcement, and output parsing that does not exist in any library.
02
19 scrapers and data sources with data quality enforcement
Orchestrating 19 Apify actors and data sources across a research job requires actor initialization, monitoring, retry logic, and result collection. Beyond orchestration: raw scraped data is noisy. Deduplication, quality filtering, and normalization must run before the data reaches the XML structuring layer. A single bad scrape cannot corrupt the dossier.
03
Tiered data integrity — Tier 1 is never overridden
The three-tier data model is programmatically enforced. When LinkedIn scraping says a person started a role in 2023, no AI web search can override that with a different date. The XML structuring layer tags each data point with its source tier before Claude sees it. Claude's instructions include explicit rules about tier precedence. Enforcing this without human review on every output required careful prompt engineering and output validation.
04
Lane-based intelligence for multi-channel companies
One project can have multiple revenue channels — enterprise software, professional services, and a SaaS product, for example — each with separate ICPs, personas, and value propositions. The segmentation engine must correctly assign each researched company to its relevant lane and apply the right scoring criteria and research prompts per lane without mixing them.
05
Self-configuring setup — zero manual input
The setup wizard researches your company automatically. It identifies your ICP, maps personas, extracts pain points your product solves, generates differentiators, and writes value propositions — all from your website URL. The output configures the entire research pipeline without a consultant or onboarding session. Building this required an LLM chain that produces structured, validated configuration data rather than freeform text, with fallback handling for companies with minimal web presence.
IPS-Builds.com

Have a data problem worth
building a real pipeline around?

Lead Detective started as a question: what if a human analyst's 5-day research workflow could run in 8 minutes? The answer required 19 scrapers and data sources, a new concurrency model, and Claude running as a subprocess. That is the kind of engineering IPS does. If the problem is hard enough, we want to hear it.