KAYANAK

Overview

Built for: Car buyers and sellers in Riyadh who speak Arabic or English, plus the KAYANAK sales team, managers, partner dealers and marketers who promote cars for commission.

Main features

• Showroom with installment pricing

  Filter by make, body type, model and price, switch between grid and list views, and see the cash price next to an estimated monthly installment on every card and detail page.

• Financing calculator and eligibility check

  Customers estimate a no-down-payment plan over one to five years with a final balloon payment, then request financing from the car page. Sales staff get an internal eligibility check that applies several Saudi bank policies and returns a green, yellow or red result.

• Smart bilingual search

  A deterministic parser normalizes Arabic text, maps Arabic make and model names, corrects typos and pulls out make, model, year, body, fuel, color and price. A query like "family suv under 100k" becomes a filtered list of SUVs under 100,000 SAR.

• AI sales assistant

  An LLM agent with tool calling searches live stock, shows car cards, takes financing requests and steps back once a salesperson replies. It runs on WhatsApp, and the same agent drives a web chat with history and an in-app chat with voice input.

• Sales CRM

  An eight-stage Kanban lead board with saved views, notes, tasks, an interaction timeline, WhatsApp templates, CSV exports and Arabic to English message translation. Leads are assigned round-robin and reassigned automatically after 15 minutes without action.

• Multi-channel lead capture

  Leads arrive in the CRM from website inquiries, sell-car requests, WhatsApp, the AI assistant and TikTok lead forms, and each request triggers a WhatsApp confirmation to the customer.

• Marketer (affiliate) program

  Referral links with UTM tracking and a 90-day attribution window feed a conversion, commission and payout chain. Marketers apply online, track their sales in a portal and generate branded posters that show the cash price and monthly installment.

• Sell your car and dealer portal

  Sellers complete a five-step guest flow (car details, condition, photos, contact, inspection appointment) and can track their request. Reviewed cars are shown anonymously to partner dealers, who make offers through their own portal.

• On-demand car wash booking

  Services priced by vehicle type and city, time slots with capacity limits, a Mapbox location picker with service-area polygons, and a booking state machine with audited overrides.

• Quotations, content and push notifications

  Staff print A4 quotations for customers, schedule hero slides and banners targeted by make or body type, and reach customers and marketers through Firebase push notifications with an in-app inbox.

Development and architecture

Modular monolith with events

Controllers stay thin, services own the logic and in-process events decouple commissions, CRM and notifications. I kept it deliberately single-instance and documented the in-memory rate limits and caches as an explicit constraint.

Authentication and access control

Argon2id password hashing, 15-minute access tokens and hashed refresh tokens with family rotation and reuse detection. Ten roles map to permissions, and a test fails if any admin route is missing a role or permission check.

Integrity enforced in the schema

Unique constraints make a second commission for the same conversion impossible, and financial totals are computed from aggregates so they cannot drift from the underlying records.

Guarded AI agent

Building on a teammate's first version, I hardened the agent for production. The LLM layer sits behind a provider abstraction (OpenAI, Anthropic or a fake driver, so tests run without API keys). An output guard rejects any price, percentage or link that did not come from that turn's tool results, and daily caps plus an admin kill switch limit cost and risk.

Semantic search foundation

Each car gets a search document and a 768-dimension Gemini embedding stored in pgvector. A background worker re-embeds changed cars and reuses vectors for identical texts. I built it as the base for hybrid AI search alongside the deterministic parser.

Arabic-first, right-to-left UI

Locale lives in the URL (/ar and /en), layouts use logical CSS properties so they flip cleanly between right-to-left and left-to-right, and each locale has a single serializable dictionary.

Integrations

WhatsApp Business messaging through Hatif, SMS one-time passwords, TikTok lead forms and pixel events, Mapbox maps and geocoding, Firebase Cloud Messaging, Cloudflare R2 storage and SMTP email.

Technology

• NestJS
• TypeScript
• Next.js
• React
• PostgreSQL
• Prisma
• Flutter
• Tailwind CSS
• pgvector
• Google Gemini
• WhatsApp API
• Firebase

Results

To build the first working version

7 daysTo build the first working versionShowroom, wash booking, sell-your-car, dealer leads, affiliates, admin and app

Cars listed in the live showroom

1,841Cars listed in the live showroomPublic listing count, November 2026

To embed every listed car for semantic search

39 sTo embed every listed car for semantic searchAbout USD 0.01 in embedding costs

Pull requests I merged in the first 27 days

140Pull requests I merged in the first 27 days