LeadForge AI
AI-powered B2B sales intelligence platform built with multi-agent workflows.
LeadForge AI automates the full B2B sales cycle: company research, lead qualification and personalized outreach — powered by AI agents, RAG and structured outputs.
Project by Alchemia StudioAI SaaS Platform
B2B sales teams
Research → qualification → outreach pipeline
01Project overview
LeadForge AI is a sales intelligence platform built as an orchestrated multi-agent system — one workflow that covers the full B2B sales cycle.
What was built
A multi-agent AI platform: company research, lead qualification and personalized outreach generation in a single workflow.
Who it is for
B2B sales teams and business developers who spend significant time on manual research, data analysis and personalization.
Main objective
Automate the full sales cycle — from finding companies to ready personalized proposals — with specialized AI agents.
Technology focus
Multi-agent LLM architecture with RAG-grounded context, structured outputs and human-in-the-loop validation.
02The challenge
Before designing the system, we mapped the real problem of B2B sales teams — no assumption, no generic template.
The problem
- Research is manual: finding and analyzing companies takes hours per lead.
- Data is scattered across multiple tools with no unified view of an account.
- Lead qualification depends on subjective judgment instead of structured criteria.
- Outreach is generic and does not reference the specific context of each company.
Constraints
- No single public dataset is complete enough to be the only source of company data.
- Sales teams need transparent and verifiable AI output — not a black box.
- The system must fit existing workflows without forcing a full process rebuild.
Why a traditional approach does not work
- A generic LLM chatbot cannot execute a multi-step workflow: research → qualification → message.
- Rule-based automation breaks on unstructured web data and free-form research.
- A CRM with bolted-on AI features does not automate the research and personalization loop.
03Solution
LeadForge AI is an orchestrated multi-agent system. Each agent owns one stage of the sales cycle and produces typed, verifiable results.
AI Company Research
Research agent enriches companies with market data from web sources.
ICP & Persona Modeling
Agents define the ideal customer profile and buyer personas from data.
Lead Qualification
Transparent scoring engine with structured, explainable criteria.
Personalized Outreach
Content agent generates messages grounded in company context via RAG.
Multi-Agent Orchestration
Workflow engine coordinates agents stage by stage with typed contracts.
Sales Intelligence
Analytics and recommendations on accounts, pipelines and performance.
04Workflow
A look inside the agent execution flow: each step returns structured results that the next agent consumes.
Business goal input
01Define target market, ICP and campaign goals.
Research Agents
02Agents gather and analyze company data from web sources.
Company Intelligence Layer
03Structured profiles, signals and context are built for each account.
ICP Analysis
04Accounts are matched against ideal customer criteria.
Qualification Agent
05Leads are scored with transparent, explainable rules.
Personalized Outreach
06Messages are generated from the company context via RAG.
Human Review
07Sales teams review and approve before anything is sent.
05Product strategy
The product strategy translates the problem into a concrete product: agents do the research, salespeople make the decisions.
Product vision
A sales intelligence platform where AI agents handle the research and people control the decisions.
User flow
Define ICP → add companies → agents research → qualification → human review → personalized outreach ready for approval.
Core features
Company research, qualification scoring, persona-based messaging and workflow automation.
AI opportunities
Automate research, standardize qualification, generate context-aware outreach and learn from outcomes.
06AI architecture
The system is designed as a five-stage pipeline. Every layer has a clear role, a concrete responsibility and a defined technology focus.
Input
01Company lists, CRM data and public web sources enter the pipeline.
APIs, web research, CRM exportsAI Processing
02Data is normalized, enriched and converted into structured features.
ETL, embeddings, structured parsingAgents / RAG / LLM layer
03Specialized agents process data with RAG-grounded context.
LLM, RAG, vector searchValidation
04Output quality is checked, guardrails applied and critical steps reviewed.
Structured output schemas, review UIOutput
05Qualified leads, scoring and personalized outreach ready for the team.
APIs, message templates, dashboard07Engineering decisions
Why this architecture? Every key decision answers a concrete requirement of the product.
Why AI Agents
Each stage of the sales cycle needs a different role, toolset and context. Agents encapsulate responsibilities and are orchestrated independently.
Why RAG
Outreach quality depends on company-specific context. RAG grounds generation in retrieved data instead of generic model knowledge.
Why structured outputs
Downstream automation — scoring, storage, sending — requires typed, machine-readable results, not free-form text.
Why human-in-the-loop
Outreach touches customers. Critical steps are reviewed and approved by people before anything is executed.
Why this stack
TypeScript end-to-end, PostgreSQL for state, LLM APIs for intelligence — fast to build, easy to maintain, production-ready.
08Technology stack
Four layers: model providers, retrieval, the application itself and production infrastructure.
AI Models
Retrieval / RAG
Application
Infrastructure
09Implementation
How the system is engineered under the hood — real technical detail without proprietary specifics.
Architecture patterns
Orchestrated multi-agent workflow with typed contracts between stages and isolated agent responsibilities.
Integrations
Web research sources, CRM data import and structured export for outreach tools.
Data flow
Company list → research enrichment → qualification scoring → content generation → human review.
Deployment approach
Containerized service, database-backed state and environment-based configuration for repeatable deployments.
10Results
What the project delivered — real capabilities and business effects, without unverified numbers.
Built capabilities
Research, enrichment and analysis run as one coordinated pipeline.
Accounts are stored as typed, searchable profiles.
Leads are scored with transparent, explainable criteria.
Messages are grounded in company context via RAG.
Business impact
Sales teams spend less time on data gathering and drafting.
Every account is evaluated against the same structured rules.
The pipeline can be reused for new campaigns and markets.
11Principles
The principles behind the system — how Alchemia Studio designs, builds and ships AI products.
AI-first architecture
The system is designed around AI capabilities from day one, not bolted on afterwards.
Human-in-the-loop validation
Every critical AI output passes through a human check before it reaches the customer.
Structured outputs
Agents return typed, machine-readable results that downstream systems can trust.
Evaluation-driven development
Model behavior is measured and compared before changes are shipped.
Production thinking
Every component is built with monitoring, reliability and maintenance in mind.
Related resources
Deep dives into the engineering behind systems like this one.
12Related products
Related products
More products built by the studio with the same engineering approach.