Service — Gen AI development
AI chatbots and automation that know your business.
I build custom AI chatbots, WhatsApp chatbots and automation flows that answer from your own documents and data using retrieval-augmented generation (RAG). They cite their sources, hand over to your team when they should, and plug into the website and tools you already use.
What I build
- 01
Custom AI chatbot for your website
A chat assistant on your site that answers product, pricing and support questions from your own content, captures leads and hands over to a person when it should.
- 02
WhatsApp AI chatbot
The same assistant on WhatsApp through the official WhatsApp Business Platform: order status, bookings, FAQs and lead qualification where your customers already are.
- 03
RAG on your documents and database
Retrieval-augmented generation over PDFs, help docs, spreadsheets, Notion or your PostgreSQL data, so answers come from your sources with links back to them.
- 04
AI automation flows
Workflows that read, sort and act: triage inbound email and forms, draft replies, enrich leads, update your CRM and post summaries to Slack, with n8n or custom code.
How RAG works
What is a RAG chatbot?
A RAG chatbot is an AI assistant that answers from your own content instead of from what a language model happened to learn on the internet. RAG stands for retrieval-augmented generation: before the model writes a reply, the system searches your documents, help pages, product catalogue or database for the passages that match the question, and hands only those passages to the model as context. The result is an answer that reflects your current prices, policies and products, links to the source it came from, and stays up to date the moment you edit a document, with no retraining. It is also the cheapest and safest way to put a large language model in front of customers, because you control exactly what it can draw on.
- Your documents and data are split into small passages and stored with a search index (usually PostgreSQL with pgvector).
- When someone asks a question, the system finds the passages that best match it, by meaning and by keywords.
- Those passages go to the language model together with the question and clear rules: answer only from this, cite the source, say when you don't know.
- The answer comes back with links to the sources, and every conversation is logged so gaps in your content are easy to spot and fix.
Use cases
Where AI chatbots pay off
- Clinics and service businesses
- Answer timing, pricing and service questions, book appointments and send reminders on WhatsApp.
- E-commerce stores
- Order status, returns and product questions answered from your catalogue and policies, around the clock.
- Real estate and education
- Qualify enquiries with a few questions, share the right brochure and pass warm leads to your team.
- Internal teams
- A private assistant over HR policies, SOPs and project docs, so staff stop searching shared drives.
- SaaS and support teams
- Deflect repeat tickets with answers grounded in your help centre, and escalate the rest with context.
Tech stack
The tools behind it
| Layer | What I use |
|---|---|
| Language models | OpenAI, Anthropic Claude, Google Gemini, or open models when data must stay in-house |
| Retrieval | PostgreSQL + pgvector, hybrid keyword and semantic search, reranking |
| Application | Next.js, React, Node.js, TypeScript, Vercel AI SDK |
| Channels | Website chat widget, WhatsApp Business Platform (Cloud API), Slack, email |
| Automation | n8n, webhooks, scheduled jobs, CRM and Google Workspace integrations |
| Quality | Test question sets, answer and citation checks, conversation logs and usage dashboards |
Safe by design
How I keep answers accurate and data safe
- Answers only from your approved sources, with citations, and a clear “I don't know” instead of a guess.
- Handover to a human on WhatsApp, email or your helpdesk when the bot is unsure or the customer asks.
- Prompt-injection defences: the model never gets more access than the task needs, and risky actions need confirmation.
- Your data is not used to train public models; I pick providers and settings that keep it that way and document them for you.
- Personal data handled with India's DPDP Act in mind: consent, minimal collection and a way to delete it.
More on this in prompt injection and how to defend against it.
Process
Discovery and data check
A call about the questions you want answered and the jobs you want automated, and a look at the documents and data you have. You get a written proposal with scope, timeline and a fixed cost.
Prototype on your real data
A working prototype, usually within one to two weeks, tested against a list of real customer questions, so you judge answer quality before the full build.
Build and integrate
The production version: your website or WhatsApp number, admin tools to update content, human handover, analytics and the integrations the flow needs.
Launch, measure, improve
Go live, review conversation logs together, fill content gaps and tune the prompts. I stay around for fixes and the next automation.
FAQ
Questions about AI chatbot development
What is the difference between RAG and fine-tuning?
RAG looks up your documents at the moment a question is asked and gives the relevant passages to the model, so answers stay current as soon as you update a document and can cite their source. Fine-tuning changes the model itself, which is slower and costlier to update and does not add citations. For business chatbots, RAG is the right starting point almost every time.
How long does it take to build an AI chatbot?
A focused prototype on your real content usually takes one to two weeks. A production website or WhatsApp chatbot with handover, admin tools and integrations typically takes three to six weeks, depending on how many data sources and systems it connects to.
What does a WhatsApp AI chatbot need?
A WhatsApp Business Platform account (the official API, not the regular WhatsApp Business app), a verified Meta business and a phone number that is not registered on WhatsApp elsewhere. I help set this up directly with Meta's Cloud API or through a provider you already use.
What are the running costs after launch?
Three things: the AI model's usage fees, which grow with the number of conversations; Meta's WhatsApp message charges if you use WhatsApp; and hosting. These are billed to your own accounts, so there is no markup, and I estimate them for your volume in the proposal.
Can the chatbot reply in Hindi, Gujarati or other Indian languages?
Yes. Current models understand and reply in Hindi, Gujarati, Marathi, Tamil and many other languages, and can answer in the customer's language even when your documents are in English. I test the languages you need during the prototype.
What happens when the bot doesn't know the answer?
It says so, offers to connect the customer to a person and logs the question. Those logs show exactly which answers are missing from your content, which is usually the fastest way to improve the bot.