Shubham Sinha

What I have built

Kerala Ayurveda runs Ayurveda clinics and online stores in India and the USA. These are company tools, so I describe what they do and show no screens.

  1. WhatsApp support bot

    Live since Aug 2026As Product Analyst

    65% fewer tickets

    80 to 90% closed without an agent

    I sorted what customers ask and when a person must take over. Then I designed the flow and built it on LimeChat, a WhatsApp chatbot platform.

    See how I built the bot: WhatsApp support bot
  2. EKAM, our clinic system

    Live at 2 sitesAs Associate Product Manager

    600+ consultations

    400+ therapies recorded

    Our clinics ran on paper. I sat with doctors and clinic managers, wrote the specs, tested the new app against the old one and trained the staff. The other 19 clinics are planned to go live by the end of December.

    See how EKAM went live: EKAM, our clinic system

Built with AI. I write the spec and check the work. AI coding tools write the code.

  1. AI search

    In testing (US store)As Associate Product Manager

    Our paid search app found nothing for “pet dard” (Hinglish for stomach ache). I specified a search that understands Hinglish, Hindi and Malayalam, and checked it against 50,000+ test searches. It returns more relevant products than the paid app.

    See how the search works: AI search
  2. Customer assistant

    In testingAs Associate Product Manager

    The bot on our website answers only inside fixed boxes. I defined an assistant that answers in the customer’s language and knows when to send someone to a doctor.

    See how the assistant decides: Customer assistant
  3. Lead Desk

    In stagingAs Associate Product Manager

    Our clinic enquiries lived in a sheet. I own Lead Desk, which records every call, its outcome and every stage change. A manager can see where a lead went cold.

    See how Lead Desk works: Lead Desk

How I build AI products

I start from the customer’s problem, not the model. In the AI search, typed results come straight from our catalogue. AI only reads a search when the need is unclear, and it has four seconds to answer. In a sample of 200 real searches, 141 never needed AI. Then I test before any customer sees it.

  • Test before launch

    The AI search was checked against 50,000+ test searches, from real search history and ones I wrote. The assistant has about 5,000 test examples, plus about 100 scripted turns in nine languages, including urgent symptoms and attempts to trick it.

  • Stop it making things up

    The search can only show products from our own catalogue. The assistant answers only from our 248 products and 22 approved answers, and code checks every reply for price, product name, link and language. A reply that fails is dropped.

  • Fixed rules for health

    An emergency such as "chest pain" gets emergency numbers and no products. A mental health search gets a helpline and a doctor link. A cholesterol question gets a link to our online consultation.

  • People in the loop

    Before customers use the assistant: a clinician review, a native-speaker review of each language, a legal review of product claims and a security check. The brand team can pin, move or hide search results without code.

  • Measure against what we had

    I built a scoring sheet that runs the new search and the paid app on the same searches. The new search returns more relevant products.

Also built

  • AI analytics system

    In useAs Product Analyst

    It flags sudden changes in revenue, order value and order count. I wrote the logic that finds the likely cause: price per item, items per order or units per item.

  • Dashboards

    In useAs Product Analyst

    Actual against plan for India and the US, return on ad spend by product and campaign, and affiliates. I wrote the SQL. They became the growth team’s source of truth.

  • Vaidyam

    LiveAs Associate Product Manager

    Our quarterly journal, as a website. I owned it end to end, through launch.

    Visit the Vaidyam site