LeadFlow AI

Problem Statement

B2B sales teams spend excessive hours manually vetting leads, triaging duplicate records, and updating disparate systems. Inbound prospects slip through the cracks without immediate qualification or structured CRM synchronization.

Dataset

Inbound prospect payloads and enterprise contact inquiries processed through normalized email indexing to ensure complete duplicate prevention across CRM records.

Architecture

Built on Next.js 15 App Router with server actions and route handlers. Features Groq Llama 3.3 for structured JSON scoring, a direct backend webhook dispatch to Twenty CRM for Person & Company upserts, and outbound Zapier catch hooks with cryptographically unique event IDs and 3-path branching.

Model Selection

Groq-hosted Llama 3.3 70B was chosen for its ultra-low latency inference, enabling real-time scoring (0–100) coupled with deterministic priority thresholds (High 80-100, Medium 50-79, Low 0-49) to ensure 100% predictable downstream routing.

Training Process

Engineered high-accuracy structured prompt contracts with JSON schema validation, deterministic fallback logic, and multi-channel routing templates for automated Gmail alerts and scheduling.

Evaluation Metrics

Results

A live sales automation platform featuring 10 operational KPIs, interactive Kanban and table pipeline views, audit logs, and verified end-to-end CRM dispatch and Zapier multi-channel routing.

Key Learnings

• 1Combining deterministic priority rules with LLM reasoning delivers robust, hallucination-free sales qualification.
• 2Normalizing emails and blocking duplicates at ingestion avoids downstream CRM pollution and redundant webhooks.
• 3Warm editorial styling with 13.8:1 contrast dramatically improves usability in high-density CRM dashboards compared to harsh clinical interfaces.