AI engineer who builds AI features into real software and tests what the AI says before customers see it
I build AI-powered software for businesses: knowledge assistants grounded in your own documents, chatbots and lead qualifiers, agent workflows and integrations with your CRM, e-mail and WhatsApp. I work with Claude, ChatGPT and Gemini models through their APIs. I do not train new models. What I add is engineering and testing, so answers are checked before customers see them.
What you get
- Working AI feature
- an assistant, qualifier or agent workflow for the agreed use case.
- Knowledge base
- your documents cleaned, structured and indexed for retrieval.
- Test set
- questions and correct answers you keep, with a record of every run.
- Integrations
- CRM, e-mail, WhatsApp or database connections with minimal permissions.
- Controls
- rate limits, spending caps, fallbacks and a human handover route.
- Conversation log
- a private record of what was asked and answered, without needless personal data.
- Admin switch
- one place to pause the assistant or change its approved content.
- Handover
- source code, prompts, settings list and a recorded walkthrough.
What an AI engineer does in a business, and what I do not do
Most business value from AI today does not come from training new models. It comes from taking a strong existing model and wiring it into your documents, your tools and your rules, with limits around what it may say and do. That is the work I do: choosing the model, giving it the right information at the right moment, connecting it to your systems and proving that it behaves.
I should be clear about the edges. I am a software engineer who builds with AI, not a research scientist. I do not train or fine-tune foundation models, and I do not sell dashboards that claim to read your customers' minds. If a problem is better solved by a simple rule, a search box or a form, I will say so, because a plain solution is cheaper to run and easier to trust.
Knowledge assistants built on your own content (RAG)
Retrieval-augmented generation, usually called RAG, lets an assistant answer from your documents instead of from what a model remembers from the internet. Your policies, product sheets, price rules, FAQs and past answers are split into passages, indexed, and the right passages are fetched for each question. The model writes the answer from those passages and can show where it came from.
The quality of such a system is decided by unglamorous work: cleaning and structuring the source documents, choosing how to split them, ranking what is retrieved, and saying clearly when the documents do not contain an answer. A good assistant says that it does not know and hands the question to a person; a bad one invents. I build for the first behaviour and test for it.
Chatbots, lead qualifiers and agent workflows
For customer-facing chat I build assistants that answer from approved content, ask the few questions that qualify an enquiry, book a call or collect details and pass the conversation, with its history, to your team through e-mail, your CRM or WhatsApp. The assistant on this website is an example of the pattern, trained on the business's own knowledge.
For back-office work I build agent workflows, where a model takes several steps with tools: reading an incoming enquiry, looking up a record, drafting a reply, updating the CRM and flagging exceptions for approval. I have run a Node.js platform with more than twenty AI agents as background services, covering content, QA checks, reporting and lifecycle tasks, so I know how agents drift, loop and fail, and how to keep them in bounds.
Testing AI output: the part that decides whether it is safe to launch
A conventional program gives the same output for the same input; a language model does not, so testing has to be designed differently. Before launch I assemble a set of real questions, including awkward, ambiguous, hostile and out-of-scope ones, and the answers a correct system should give. Every change to the prompt, the model or the documents is run against that set, and failures are read by a person rather than only scored by another model.
I look for specific faults: invented facts, outdated prices, answers that ignore the documents, leaked internal instructions, rude or off-brand tone, replies to questions the bot should refuse, and instructions hidden in a user's message that try to redirect it. The test set is yours, so it keeps protecting the system long after I leave.
Privacy, cost control and keeping a human in the loop
AI features handle customer text, so privacy is designed in. I send only the information a task needs, keep personal details out of logs, choose providers whose data terms match your obligations, and mark where laws such as GDPR, UK GDPR or India's DPDP Act 2023 may apply, leaving final advice to your lawyer. Secrets stay on the server, never in the browser.
I also build the brakes: limits on how many requests one visitor may make, a monthly spending cap per provider, a fallback message when a provider is down, and a clear route to a human for anything sensitive, such as complaints, medical questions or payment disputes. You can see what the assistant said, to whom and why, and switch it off in one place.
Connecting AI to the tools you already use
An assistant that cannot act is only a talking FAQ. I connect AI features to the systems where work happens through their official APIs, MCP or OpenAPI definitions: your CRM, e-mail, calendar, WhatsApp through providers such as WATI, spreadsheets, ad platforms, analytics and your own database. Each connection is given the least access it needs, and actions that change data can require approval first.
My background in performance marketing helps here: lead qualification, follow-up timing, tracking and attribution are things I have run for twelve years, so the AI is attached to a process I understand rather than to a demo.
How the work runs
- 01
Pick the use case
A call to choose one task where AI saves real time or wins real leads, with a measure of success agreed in advance.
- 02
Gather and clean content
Collect the documents, FAQs and rules the assistant may use, remove outdated material and mark what it must never say.
- 03
Build a prototype
A working assistant or workflow on a private link, built on the model and tools that suit the task and the data rules.
- 04
Create the test set
Real, tricky and hostile questions with correct answers, run against every change and reviewed by a person.
- 05
Connect and add controls
Integrations with your tools, spending and rate limits, logging, approval steps and a clear handover to people.
- 06
Launch, watch and tune
A limited launch, daily review of real conversations for the first weeks, then fixes to content and prompts.
Tools and proof
- Tools
- Claude, ChatGPT and OpenAI APIs, Gemini, Node.js, Python, Next.js, MCP and OpenAPI, n8n and WATI
- Multi-agent system
- A Node.js platform with 20+ AI agents as background services for content, QA checks, reporting and lifecycle tasks.
- Models used daily
- Claude, ChatGPT, Gemini, NVIDIA APIs and DeepSeek through Ollama, integrated through their APIs.
- Chatbots and RAG
- Knowledge-base assistants and lead-qualifying chatbots grounded in a business's own content, like the assistant on this site.
- Engineering degree
- M.Tech in Computer Science and Engineering, with 12+ years of working with 500+ brands.
Questions about AI engineering and AI-powered software
No. I build applications on top of established models such as Claude, ChatGPT and Gemini through their APIs. The engineering is in the content, retrieval, integrations and testing, which is where most business results come from.
I restrict it to approved documents, require it to say when the answer is not there, hand uncertain questions to a person and test it with a fixed set of awkward questions after every change. This reduces invented answers; it cannot remove them entirely.
I send only what a task needs, keep secrets on the server, avoid personal details in logs and choose providers whose terms fit your obligations. I explain the trade-offs in plain words, and your lawyer should confirm the legal position.
A normal chatbot answers from what the model already knows. A RAG assistant first looks up passages from your own documents and answers from those, so it can follow your current policies and cite where an answer came from.
It can do both. It can update a CRM, book a slot or draft an e-mail, each with the least access needed, and anything that changes important data can require a person's approval before it happens.
We agree measures at the start, such as share of questions answered correctly on the test set, qualified leads passed to your team, time saved on a task or escalations to humans. Then I review real conversations and report against them.
Yes. I direct AI coding tools such as Claude, Codex and Cursor to write code in small steps, then review, run and test every change myself. I use the same tools to speed up testing, and a person reads every result.
AI engineer by country and city
- AI engineer in AhmedabadAhmedabad firms usually pay by NEFT, RTGS or cheque against a GST invoice, with UPI for smaller sums. Traders and manufacturers pay close attention to the GSTIN and invoice format, and some deduct TDS before paying a professional fee, so I suggest agreeing these points in the quote.
- AI engineer in BangaloreBangalore companies usually pay by bank transfer (NEFT or RTGS) against a GST invoice, and funded start-ups often follow a purchase-order or vendor-onboarding process with TDS deducted at source. It saves time to confirm the invoice details and payment schedule in writing at the start.
- AI engineer in ChennaiChennai firms typically pay professional fees by NEFT or RTGS, and UPI for smaller amounts. Manufacturers and exporters expect a GST invoice with the supplier's GSTIN, and some deduct TDS at source, so invoicing details are best agreed in the written quote.
- AI engineer in DelhiDelhi businesses typically pay professionals by bank transfer (NEFT or RTGS) and use UPI for smaller sums. Traders and companies expect a GST invoice with the supplier's GSTIN, and some deduct TDS before paying a technical or professional fee, so those details are best settled in the quote.
- AI engineer in GurgaonGurgaon companies, especially larger corporates, usually pay professional fees by NEFT or RTGS after vendor onboarding, against a GST invoice with the supplier's GSTIN. TDS is commonly deducted at source and payment follows an agreed credit period, so these terms are best written into the quote.
- AI engineer in HyderabadHyderabad companies mostly pay professional fees by NEFT or RTGS, with UPI for small amounts. IT and pharma buyers usually expect a GST invoice and a vendor form, and many deduct TDS before payment, so I suggest agreeing these details at the quote stage.
- AI engineer in KolkataKolkata businesses mostly pay by NEFT, RTGS or cheque to a bank account, with UPI for smaller sums. Traders and companies expect a GST invoice with the supplier's GSTIN, and some deduct TDS, so I suggest settling invoicing details in the quote.
- AI engineer in MumbaiMumbai companies commonly settle professional fees by NEFT or RTGS to a bank account, with UPI used for smaller amounts. Registered businesses expect a GST invoice, and many corporate and financial-services clients deduct TDS before paying, so it helps to agree the invoicing details in the quote.
- AI engineer in NoidaNoida firms usually pay professional fees by NEFT or RTGS, and by UPI for smaller amounts, against a GST invoice. Companies with vendor processes may pay after a credit period and deduct TDS, so the invoicing terms are best written into the quote.
- AI engineer in PunePune businesses commonly pay by NEFT or RTGS against a GST invoice, and manufacturing and IT buyers often work through a vendor registration and purchase order, with TDS deducted at source. Agreeing the invoice details and the payment schedule in the quote avoids delays later.
Related services
- AI chatbot developerAI chatbot and RAG assistant development: answers from your own documents, lead qualification, WhatsApp and website chat, CRM hand-off and testing for errors.
- Workflow automation specialistWorkflow automation for lead follow-up, CRM updates, reporting and WhatsApp and e-mail messages, built with Zapier, Pabbly, WATI and custom Node.js code.
- Custom software developerCustom software for business: CRMs, back-end and API workflows, integrations and reporting tools. Scoped in writing, tested, documented and handed over.
- Web application developerCustom web applications built with Next.js, React and Node.js: client portals, dashboards, booking and internal tools, tested and handed over with the code.
- CRM developerCustom CRM development: lead capture, pipelines, roles, WhatsApp and e-mail follow-up, reports and multi-tenant dashboards. Built remotely, owned by you.
Need AI engineering and AI-powered software?
Tell me the goal and the market; you get a plain answer and a fixed quote, usually within one working day.
Fastest reply on WhatsApp — usually within a couple of hours.