Lead Generator Proposal Engine

Overview

The Lead Generator Proposal Engine is an internal AI-powered sales tool for a commercial energy management consultancy. Given a prospect's company details — or a photo of their utility bill — the system runs a multi-agent pipeline that researches real energy-market data and generates a fully customized savings proposal in under 60 seconds.

Problem

Energy consultants spend hours per prospect manually researching state electricity rates, solar incentives, and compliance requirements before hand-assembling a savings proposal. The process is slow, inconsistent, and hard to scale across a sales team.

Solution

A four-phase, multi-agent pipeline that turns a prospect profile into an adjudicated, financially-modeled proposal — plus a ready-to-send client email:

• Bill intake: GPT-4o vision extracts spend, usage, and location straight from an uploaded utility bill
• Grounded research: every run is anchored to real EIA electricity rates and NREL PVWatts solar data, not model guesses
• Isolated specialists: parallel agents cover usage reduction, procurement, alternative generation, and compliance
• Synthesis + financials: a final analysis layer produces a five-year savings projection with paybacks
• Real-time UX: agent status streams to the frontend live via Server-Sent Events

Architecture: Multi-Agent Pipeline

The orchestrator runs agents across four sequential phases:

Phase 1 — Research (serial, blocking). Fetches the real commercial electricity rate for the prospect's state from a 50-state EIA 2024 table (with live EIA API fallback), queries a RAG store for the top-3 most similar past proposals by cosine similarity, and produces a structured research brief that every downstream agent shares.

Phase 2 — Specialist Agents (parallel). Four agents run concurrently, each receiving the research brief but operating in isolation so bias can't propagate between them:

Each recommendation is a structured item with a confidence score and a source.

Phase 3 — Financial Analysis (serial). The only agent with visibility into all specialist outputs, acting as a synthesis layer — it produces a baseline, a savings breakdown by category, blended payback, and a five-year cumulative projection.

Phase 4 — Email Drafting (serial). Generates a personalized subject and body ready for the consultant to send.

Fault tolerance. Every agent is wrapped in a safeRun() guard — if one fails or times out, the orchestrator logs it, emits an error event to the UI, and substitutes a sensible default so the pipeline always completes.

Tech Stack

• Frontend: React 18 + Vite + Tailwind CSS
• Backend: Express (ESM), Server-Sent Events for live agent status
• AI: OpenAI SDK — GPT-5.5 for complex agents, GPT-4o-mini for specialists, GPT-4o vision for bill extraction, text-embedding-3-small for RAG
• Database: SQLite (better-sqlite3) for proposals and RAG embeddings
• Email: Nodemailer over SMTP
• Grounding data: EIA state electricity rates, NREL PVWatts solar production

Key Design Decisions

• Specialist agents are isolated — they read only the shared research brief, never each other's outputs, preventing bias propagation.
• Financial analysis runs last as the single synthesis layer with full visibility.
• RAG is fire-and-forget — indexing and retrieval failures are swallowed so proposal generation always degrades gracefully.
• SSE streaming keeps the sales rep informed as each agent completes, rather than staring at a spinner.

Results

• Generates a complete, grounded savings proposal in under 60 seconds
• Every recommendation is traceable to a confidence score and a real data source
• Scales consistent proposal quality across an entire sales team