Projects / Scarlett
Live on this site
Conversational Sales Intelligence

The agent that qualified you when you opened this site.

Sector · AI / B2B Sales
Status · Production
Owner · Hash Include

Scarlett is a conversational sales agent embedded across hashinclu.de. She qualifies visitors, scores intent on five dimensions, captures contact details, and auto-generates a client-ready proposal — all before a human touches the conversation.

5D
BANT+F qualification dimensions, scored 0–100
2-tier
Rule-based + LLM qualification engine
36
Curated knowledge chunks with TF-IDF retrieval
AR / EN
Auto-detected, fully separated language modes
0
Vector DB cost. Zero embedding API calls.
The problem
Most B2B websites lose the lead at the form. Or worse — at the chatbot.
The status quo

Static contact forms with five fields, no context, no qualification — every submission gets the same generic auto-reply.

Off-the-shelf chatbots that announce themselves as bots, dump information, and cannot tell a buyer from a job seeker.

No proposal, no follow-up SLA — qualified leads sit in an inbox until a human gets to them, hours or days later.

Sales teams flying blind — no scoring, no signals, no record of what the prospect actually said before the first call.

With Scarlett

A real conversation — Scarlett presents as a team member, asks one question at a time, and probes for the real problem behind the request.

Lead captured naturally — name, mobile, email pulled mid-conversation in the user's own language, English or Gulf Arabic.

Auto-generated proposal — once Scarlett has enough context, a branded DOCX proposal is generated and emailed to the prospect within seconds.

Sales armed with context — every lead arrives with a tier, a score, BANT+F signals, industry tags, and the full transcript.

The opening move
She doesn't take requests at face value. She probes.

Most chatbots answer questions. Scarlett asks them. When a visitor describes a solution — "I want to build an app that scans license plates" — Scarlett's first move is to surface the underlying problem: tracking? Access management? Something else entirely?

This is consultative selling encoded into a system prompt. The agent stays curious, never pitches unprompted, never dumps information. Two to three sentences per response. One question at a time.

By exchange three, contact details have been captured naturally. By exchange six, a proposal is in the prospect's inbox. The visitor never feels qualified. They feel understood.

S
Scarlett · Hashinclude
● ONLINE
Hey! I'm Scarlett from Hashinclude. What are you working on?
I need an app that scans license plates
Got it — what happens after you capture the plate? Tracking, access management?
Access control for our compound — 3 gates, ~400 vehicles
Makes sense. Sounds like something we can definitely build. Let me grab your details so the right person can reach out:

NAME:
MOBILE:
EMAIL:
What we built
Eight components. One agent. Built from scratch, no chatbot platform.

Persona Engine

System prompt scaffolding that maintains Scarlett's identity, deflects technology questions, and never breaks character — even under direct probing.

Two-Tier Qualifier

Rule-based regex scoring on every message + LLM-based refinement once email is captured. Falls back gracefully if the LLM fails.

Contact Extractor

Multi-pattern regex extraction for name, mobile, email, and company — pulls structured data from natural conversation in real time.

TF-IDF RAG

36 curated knowledge chunks scored by term frequency and inverse document frequency. No vector DB. No embedding API. Zero retrieval cost.

Proposal Pipeline

Token-triggered LLM extraction → structured data → branded DOCX generation → dual email delivery to prospect and internal team.

Bilingual Mode

Auto-detects Arabic vs. English and switches entire response language. Gulf Arabic (خليجي) for Arabic input. Never mixes languages.

Lead Dashboard

Live dashboard with tier badges, BANT+F signal indicators, budget mention quotes, industry tags, full transcripts, and re-qualification.

Session Persistence

Two-tier storage: sessionStorage for in-tab continuity, localStorage for returning users. Cloud Storage in production, filesystem in dev.

The hard part
A chatbot that captures leads is easy. A chatbot that scores them is the system.

Every message a visitor sends is scored on five dimensions — Budget, Authority, Need, Timeline, Fit. Each worth twenty points. Maximum one hundred. This runs on every message, with zero LLM cost.

Rule-based regex catches the obvious signals — currency mentions, decision-maker language, urgency phrasing, problem clarity, sector alignment. The LLM tier kicks in once an email is captured, refining the score with full conversational context at temperature 0.2 for consistency.

Negative signals are detected too. Job seekers get auto-classified to score zero. Students, free-help seekers, competitor probes, and price shoppers each take a fifteen-point penalty. The sales team sees only what's worth seeing.

B
Budget

Currency mentions, pricing, investment language, exact amounts

/ 20
A
Authority

Decision-maker language: CTO, founder, "my team", "I lead"

/ 20
N
Need

Clear problem: "we need", "currently struggling", "rebuild"

/ 20
T
Timeline

Urgency: "by Q3", "ASAP", "next month", "ready to start"

/ 20
F
Fit

Sector alignment: government, regulated, AI, enterprise

/ 20
Hot
70 – 100
Warm
40 – 69
Cold
20 – 39
Unqual
0 – 19
The closer
Most chat tools stop at "we'll be in touch." Scarlett sends the proposal.
01

Trigger Detection

Scarlett emits a hidden token once name, organisation, email, and project clarity are present. Token is stripped before display.

Single-shot per session
02

LLM Extraction

Second LLM call extracts structured data: client, organisation, sector, problem, approach, deliverables, timeline, team, next steps.

Temp 0.3 · 2048 tokens
03

Case Study Match

Three relevant case studies selected by detected sector — government work for ministries, FinTech for banks, InsurTech for brokers.

Sector-aware
04

DOCX Generation

Branded Word document assembled — exec summary, approach, deliverables, timeline, three case studies, next steps, pricing notes.

Hash Include template
05

Dual Delivery

Professional email to the prospect with attachment. Internal notification to sales with lead summary, BANT+F signals, and full transcript.

Gmail SMTP
What makes this hard
A chat widget is a weekend project. This is a system.

Persona Discipline

Scarlett never confirms she's AI. Never reveals the model. Never breaks character. She handles "are you a bot?" without lying and without admitting.

Two-Tier Scoring

Rule-based scoring runs on every message for instant tier assignment. LLM refinement runs once on email capture. The system stays fast, cheap, and graceful when the LLM fails.

Bilingual Without Mixing

Detect Arabic input, switch entire response to Gulf Arabic, never mix scripts mid-message — but allow proper nouns to stay in English.

Auto-Proposal Trigger

Knowing when to generate a proposal is harder than generating it. All four signals must be present, and the trigger fires exactly once per conversation.

Zero-Cost Retrieval

TF-IDF over 36 hand-curated chunks. No vector database. No embedding API. Retrieval is in-memory, deterministic, and free.

Graceful Degradation

If the LLM API is down, rule-based scoring still works. If Gmail fails, the lead is still captured. Every layer fails into a useful state.

Query + last 4 messages tokenised together for context-aware retrieval

36 chunks scored by Term Frequency × Inverse Document Frequency

Boosts applied — exact keyword matches +2, title matches +1.5

Top 6 chunks selected + company-overview always pinned

Injected as RELEVANT KNOWLEDGE into the system prompt

RAG without the bill

Vector databases are often unnecessary.

Most RAG implementations reach for a vector database, an embedding API, and a recurring monthly bill — for a knowledge base of fewer than fifty documents.

Scarlett's knowledge base is thirty-six curated chunks. At that scale, classical TF-IDF retrieval is faster, cheaper, and more deterministic than any vector approach. Same query, same answer. Every time.

This is the same engineering judgement we apply to every system we build. Use the simplest tool that solves the problem.

Scarlett is not a client project. It is our own product — and our most honest sales reference.

We built Scarlett to handle our own inbound, on our own site, against our own pipeline. Every conversation she has is a conversation our sales team would otherwise have. The fact that you are reading this page means she is doing her job.

1×
Live deployment, our own site
v2
Current generation, in production
0
External chatbot platforms used
Who this is for
If your inbound funnel ends at "thanks, we'll be in touch" — there is a better version of this.

Service businesses

Consulting firms, agencies, system integrators, law firms, accounting practices — anyone whose website inquiries currently get the same generic auto-reply.

Regulated industries

Insurance brokers, financial advisors, healthcare providers, education institutions — sectors where qualification and compliance both matter from message one.

Bilingual markets

Any UAE, GCC, or MENA business serving Arabic and English speakers in parallel — where language detection on every turn is non-negotiable.

We do not build Scarlett-like systems. We build Scarlett-grade systems.

117+
Production systems delivered
41
Platforms in active stewardship
21+
Years of engineering experience
Want one of these?

Scarlett is right there in the corner. Ask her about it.

Or skip the demo. Tell us about your inbound funnel and we will scope a deployment for you — branded as your own, trained on your knowledge base, integrated with your CRM.

S
Scarlett Hashinclude AI
Online
S
Hey! I'm Scarlett from the Hashinclude team. I'm here to understand what you're working on and connect you with the right people.

Tell me about what you're building or trying to solve.
Just now
S
Enter to send · Shift+Enter for new line
anas.exe
> _
$
Powered by Hashinclude AI
WhatsApp