Internal Engineering Demonstration
AI Lead Qualification Agent
Internal engineering demonstration. This is not a client deployment. All scenarios, figures and outcomes are illustrative and created to showcase NorthFlow Studio's technical approach.
Business challenge
- Inbound leads arrived across form, email and WhatsApp with no consistent triage.
- Sales reps spent time on low-fit leads while hot ones went cold.
- No single score to prioritise follow-up.
Objectives
- Lead classification latency: Automated classification within seconds
- Hot lead routing: Immediate routing with full context attached
- Manual triage reduction: Significant reduction in manual rep time spent on triage
Architecture
Form / Inbox
│
▼
[ Classify ] ──▶ spam? → discard
│
▼
[ Extract ] ──▶ name, need, budget, timeline
│
▼
[ Score ] ──▶ fit 0–100
│
├─ hot ──▶ notify human + create CRM deal
└─ warm ──▶ enroll in nurture sequence
Inbound lead → agent → scored + routed.
score.ts
export function score(lead: Lead): number {
let s = 0;
if (lead.budget) s += 40;
if (lead.timeline === "now") s += 30;
if (lead.fit) s += 30;
return Math.min(100, s);
}Technology
TypeScriptOpenAISupabasen8nWhatsApp Business API
Features
- Multi-channel ingestion (web form, email, WhatsApp) into a unified schema
- LLM-based entity extraction with typed JSON validation
- Rule-validated scoring engine (budget, timeline, fit) with override logging
- Automated nurture sequence enrollment for warm leads
- CRM deal creation with extracted lead context attached
Implementation
- 1Ingest leads from each channel into one normalized table.
- 2Run a classification + extraction step with a typed schema.
- 3Score fit and route hot leads to a human with context attached.
- 4Enroll warm leads in an automated nurture sequence.
Tradeoffs
Model vs rules
LLM extraction is flexible but needs validation; we keep a rules layer for hard constraints.
Latency
Real-time scoring adds ~1s; acceptable for lead routing, not for page loads.
Performance
- p95 classification under 1.5s.
- Batched nurture sends to protect deliverability.
Lessons learned
- Validation matters more than the model — bad extraction ruins routing.
- A simple score beats a clever one if sales trusts it.
Future roadmap
- Add learning from rep overrides to improve the score over time.
- Support more channels via a single ingestion adapter.
Demo video placeholder
Screenshot / GitHub placeholder
FAQ
Is this a client deployment?+
No. It is an internal engineering demonstration illustrating our approach. Figures are illustrative.
Can you build this for us?+
Yes — see AI Agents and Business Automation, or book a call.
Want a walkthrough for your system?
Book a free strategy call. We'll map your bottlenecks to the right systems and send a clear roadmap — even if we don't work together.
Usually replies within minutes. Chat on Telegram: @northflowstudio
No obligation consultationFounder-led projectsInternational clientsFast responseSecure communication