Kushal Sapkota: I build systemsthat think.And automate.

A simplified visualisation of the production workflow, which runs 98 nodes across 5 n8n workflows. Read the full case study

01Karkiz AI agent In production02KusBill In production

Production AI agent · n8n

Shipping an AI agent to a real shop.

A Nepali-language ordering agent for Karki’z cakes on WhatsApp and the website. Customers write Nepali, Roman Nepali and English, often in one message, and send voice notes, photos and payment screenshots.

In production for Karki’z Cakes

• n8n
• PostgreSQL
• OpenRouter
• 360dialog API
• Docker

98workflow nodes

2channels

3languages

5n8n workflows

Watch it run See the shop

01

The problem

Five constraints that shaped it

1. 01Customers write Nepali, Roman Nepali and English, often in one message.
2. 02Orders arrive as voice notes, photos and payment screenshots.
3. 03A non-technical owner must change behaviour without calling me.
4. 04Money is involved: the model must never decide a payment.
5. 05It must fail loudly. No silent drops.

02

The architecture

One message, seven stages

Channels

• WhatsApp · 360dialog
• Website chat

Ingest

• Webhook
• Dedupe
• Voice → speech-to-text
• Image fetch

Gate

• Chat paused for team?
• Owner settings
• Debounce

Agent

• LLM
• Memory
• 5 SQL tools
• Structured output

Guards · code

• Feature flags
• Items valid?
• Date valid?
• Fee calculation

Actions

• Reply + photos
• Payment QR
• Save order
• Telegram alert

State

• PostgreSQL
• Admin panel
• Inbox

Tools are SQL. Guards are code. The LLM sits in the middle, not on top.

03

The engineering decisions

Seven decisions that kept it reliable

1. 01

   The model proposes. Code decides.

   The owner’s feature switches are re-checked in code after the model answers, so “off” really means off. Totals and fees are computed in SQL, never by the model.

2. 02

   Structured output, not free text.

   One reply for the customer and machine-readable fields for the workflow. A parser fixes the shape before anything downstream runs, and the agent never writes to the database.

3. 03

   Tools, not memorised facts.

   Five SQL tools (cakes, occasions, delivery zones, time slots, payment methods) mean a price changed in the admin is used by the bot on the next message.

4. 04

   Reading payment proofs, carefully.

   A vision model reads the paid amount and a backup model tries if the first cannot. The result only flags the order for the owner to confirm; the AI cannot mark it paid.

5. 05

   Behaviour as data.

   The owner flips switches, edits FAQs and adds holiday notices in an admin panel, and a Postgres function composes them into the prompt.

6. 06

   Humans stay in the loop.

   One inbox for WhatsApp and web chats. A staff reply pauses the bot for that customer, and the owner is alerted on Telegram when it needs help.

7. 07

   Test on a copy, never the live shop.

   Tests run on a copy of the workflow with every send node disabled and fake phone numbers, and each deploy is diffed node by node.

04

Result and status

In production for Karki’z Cakes

What broke, and what changed

• Website chat went silent for about an hour.A regenerated workflow was left inactive.Now: check every workflow’s active flag after each deploy.
• A test left the bot stuck on “always English”.A customer writing Roman Nepali got an English reply.Now: tests restore production defaults.
• Meta rejected Nepali-script documents, twice.It happened during WhatsApp business verification.Lesson: compliance is part of the system design.

Original slides and recorded demos 16 slides

2 / 16

Architecture walkthrough Chat demo (recorded)

Recorded demonstration, not a live connection to the shop.

Read how it was built: How I built a WhatsApp ordering agent in n8n

• 

  01

  Karki’z AI ordering agent

  In production

  A Nepali-language agent on WhatsApp and website chat that takes cake orders, reads voice notes and payment screenshots, and hands over to staff when it should.

  ProblemOrders arrived as mixed-language messages, voice notes and photos, and the shop needed them captured without a model ever deciding a payment.

  • n8n
  • PostgreSQL
  • OpenRouter
  • 360dialog API
  • Docker

  Case study Live site
• 

  02

  KusBill

  In production

  A billing and ledger app for small Nepali shops: type bills like the paper pad, with balances calculated from history, Bikram Sambat and Gregorian dates, customer and supplier ledgers, and PDF, Excel and CSV reports.

  ProblemA paper billpad is fast until one slip makes every later balance wrong, and nobody can quickly say who still owes what.

  • Next.js 15
  • TypeScript
  • PostgreSQL
  • Prisma
  • Zod
  • Tailwind
  • Vitest

  Case study Code
• 03

  Affaiai

  In development

  An AI-powered WhatsApp automation platform for businesses, founded and owned by me. It answers customer questions from the business’s own FAQ, hands over to a person when it is not confident, and lets owners build their own flow in a guided editor.

  ProblemSmall businesses answer the same WhatsApp questions all day, and a bot that makes up prices or promises is worse than none.

  Early stage and still growing. See aafaiai.com for what is available today.

  • Django
  • Django REST Framework
  • Celery
  • Redis
  • PostgreSQL
  • React
  • TypeScript
  • Docker

  Live site
• 

  04

  AAFAI Video Maker

  Live · free to try

  A free tool where you type a short script, pick a character, and get a talking cartoon video with voice and music for TikTok, Reels or YouTube. A longer mode builds a cartoon of up to 60 seconds with different backgrounds, moods and sound effects.

  ProblemA short talking video in Nepali needs a voice, animation and editing, which is a lot of work for one person.

  I designed the product and built it with AI coding assistance (Claude). Credits for the open-source voices and tools are on the site.

  • Python
  • FastAPI
  • Piper TTS (Nepali voices)
  • FFmpeg
  • Pillow
  • nginx

  Try it free
• 

  05

  Nexus AI, an n8n chat studio

  Live demo

  A chat interface for n8n webhook workflows: paste a webhook URL, talk to the workflow, attach files, switch themes and export conversations. A built-in mock mode lets anyone try it without n8n.

  ProblemGives an n8n webhook workflow a proper chat front end instead of a REST client.

  • React 19
  • Vite
  • n8n webhooks
  • Browser storage

  Live demo Code
• 06

  Health Service AI

  Prototype

  A Django app where users describe symptoms, get a plain-language suggestion from a locally run LLM (Ollama) and review their past queries. A learning prototype, not medical advice.

  ProblemExplores giving plain-language health guidance with a model that runs on your own machine.

  • Django
  • Python
  • Ollama (Llama 3)
  • SQLite

  Code
• 07

  College Finder

  Prototype

  A Django REST platform for searching colleges by location, course, budget and facilities, with an Ollama LLM (Llama 3.2) that compares options. Has student, parent and admin roles, favourites and analytics.

  ProblemComparing colleges means weighing many factors at once; this narrows the list and explains the trade-offs.

  • Django
  • Django REST Framework
  • Python
  • Ollama (Llama 3.2)
  • SQLite

  Code
• 08

  Farmers AI

  Prototype

  A chat assistant for farmers: ask about fertilisers, crops and plant diseases, or upload a plant photo for a diagnosis. Built on Chainlit with local models through Ollama and LangChain.

  ProblemPuts farming answers in a chat anyone can use, including from a photo of a sick plant.

  • Python
  • Chainlit
  • LangChain
  • Ollama

  Code
• 09

  Autism Screening Assistant

  Prototype

  A screening-support prototype: an ensemble machine-learning model trained on a synthetic behavioural dataset, plus an LLM that reads free-text behaviour descriptions in a Chainlit chat. A screening aid, not a diagnostic tool.

  ProblemExplores how an LLM and a trained model can work together to flag behaviours worth discussing with a clinician.

  Trained on synthetic data. Not a diagnostic instrument.

  • Python
  • scikit-learn
  • Chainlit
  • LLM

  Code
• 10

  EduPortal

  Source only

  A Django exam-preparation portal with notes, video lectures, mock tests and subscriptions, plus a documented REST API.

  ProblemBrings exam material, practice tests and paid access into one portal.

  The old hosted demo is offline; source on GitHub.

  • Django
  • Django REST Framework
  • Python
  • JavaScript

  Code

Client websites are in the Client work section above. All repositories on GitHub