coefficient-portfolio

Personalized Outreach System

Five-agent AI pipeline that enriches prospect data from public sources, qualifies leads against an ICP framework, and generates personalized multi-channel campaigns — with messaging validation gates between every agent.

What It Does

Takes a Google Sheet of prospect names/emails, enriches each with company intelligence and individual research, scores them against a 100-point qualification framework, then designs and produces personalized outreach campaigns. Every agent output is validated against a messaging guide (>80 score required to proceed).

Built for a B2B SaaS startup targeting mid-market finance teams (~100-300 employees).

Architecture

┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│   Agent 1    │────▶│   Agent 2    │────▶│   Agent 3    │────▶│   Agent 4    │────▶│   Agent 5    │
│  Researcher  │     │   Analyst    │     │  Strategist  │     │   Creator    │     │   Editor     │
│              │     │              │     │              │     │              │     │              │
│ Enrich from  │     │ Qualify &    │     │ Design       │     │ Produce      │     │ Review &     │
│ public       │     │ score        │     │ campaign     │     │ assets       │     │ approve      │
│ sources      │     │ prospects    │     │ strategy     │     │              │     │              │
└──────┬───────┘     └──────┬───────┘     └──────┬───────┘     └──────┬───────┘     └──────┬───────┘
       │                    │                    │                    │                    │
       ▼                    ▼                    ▼                    ▼                    ▼
   Messaging            Messaging            Messaging            Messaging            Messaging
   Validation           Validation           Validation           Validation           Validation
   Gate (>80)           Gate (>80)           Gate (>80)           Gate (>80)           Gate (>80)

Pipeline Stages

Agent 1: Researcher

Enriches prospect data from public sources using a two-level architecture:

Company-level (researched once per company, shared across prospects):

• Company profile (size, industry, funding, growth stage)
• Trigger hunting (automation initiatives, migrations, leadership changes, M&A)
• Pain point discovery (manual workflows, data silos, slow reporting, access bottlenecks)
• Tech stack discovery (ERP, data warehouse, BI, CRM)
• Persona identification (champions, economic buyers, users)
• Use case inference (mapped to product’s top 8 use cases)

Individual-level (researched per prospect):

• LinkedIn profile discovery (5 search strategies with fallback)
• Career history analysis
• Pain point & expertise mining from LinkedIn activity
• Internal product usage data analysis
• Persona classification and champion potential scoring
• 3 personalization hooks per prospect

Tools: LinkedIn Scraper MCP, Firecrawl MCP, Exa MCP, Google Sheets MCP

Agent 2: Analyst

Qualifies prospects against ICP using Agent 1’s research:

100-point scoring framework: | Category | Points | What’s Measured | |———-|——–|—————-| | Company Fit | 40 | Employee count, industry, tech stack | | Trigger Quality | 20 | Leadership changes, funding, migration, M&A | | Pain Alignment | 20 | Direct quotes vs inferred pain | | Individual Fit | 20 | Persona quality + product engagement |

Score tiers: 90-100 (HOT), 80-89 (WARM), 70-79 (QUALIFIED), 60-69 (MARGINAL), <60 (DEPRIORITIZE)

Also produces: Use case assignment, benefit pillar mapping, competitive context, outreach strategy recommendation

Agent 3: Strategist (Planned)

• Designs multi-channel campaign strategy based on qualification tier
• Selects channels, sequences, timing, and messaging angles
• Maps content to funnel stage and persona

Agent 4: Creator (Planned)

• Produces actual marketing assets (emails, ads, landing pages)
• Personalizes based on Agent 1 research + Agent 2 qualification + Agent 3 strategy

Agent 5: Editor (Planned)

• Reviews all assets for quality, brand compliance, and messaging alignment
• Final approval gate before outreach execution

Key Features

• Two-level data architecture: Company intelligence researched once and shared across all prospects at that company. Prevents redundant API calls and ensures consistency.
• Messaging validation gates: Every agent transition scored against messaging guide (>80 required). Catches misalignment early before downstream agents waste work.
• Email parsing intelligence: Handles 5 email formats (firstname.lastname, firstname_lastname, f.lastname, firstname@, etc.) with varying success rates for LinkedIn discovery.
• 5 LinkedIn search strategies: Cascading fallback from email search → name+company → domain search → company people list → job posting analysis.
• Internal usage signals: Incorporates product usage data (login frequency, feature adoption) to distinguish active users from cold prospects — changes outreach from cold pitch to expansion play.

Results (Agents 1-2 Live Test)

Metric	Value
Company research completeness	90/100
Individual qualification accuracy	74/100 (QUALIFIED tier)
Messaging validation score	97/100
Data fields enriched per prospect	32 (18 company + 14 individual)
Time savings vs manual	65-95 min → 6-10 min per prospect
System completion	40% (Agents 1-2 done, 3-5 planned)

Stack

• Language: Python 3.11+
• Orchestrator: Claude Sonnet 4.5
• Data: Google Sheets (primary store) + Markdown (knowledge base)
• MCP servers: LinkedIn Scraper, Firecrawl, Exa, Google Sheets, Google Docs, Mem0
• Libraries: pandas, asyncio

Key Decisions

• Two-level over flat architecture: A flat approach re-researches the same company for every prospect. The two-level approach (Company Intelligence tab + individual rows) researches each company once, saving 15-20 min per duplicate.
• Messaging validation gates over end-of-pipeline review: Catching messaging misalignment at Agent 1 is cheaper than discovering it at Agent 5. Each gate adds ~30 seconds but prevents hours of wasted downstream work.
• 100-point scoring over binary qualify/disqualify: Granular scoring enables prioritization within qualified prospects and reveals which dimension (company fit, triggers, pain, individual) is weakest — informing outreach strategy.
• Internal usage data as signal: Discovering a prospect is an active user completely changes the outreach approach (expansion vs. cold). This single signal had the highest impact on campaign strategy.

Visual Direction

For designer: 5-agent horizontal pipeline with messaging validation gates shown as checkpoints between each agent. The two-level data architecture should be visualized as a shared “Company Intelligence” layer underneath the individual prospect flow. Show the 100-point scoring breakdown as a stacked bar chart in Agent 2’s section.