Kabir Gupta

A platform that explains lab reports via a phone call, WhatsApp, or the web. Live in 5 Indian languages and Hinglish (Hindi+English code-switched dialogue).

Built

It reads any Indian lab report, flags what is off, and informs the user when to see a doctor without a diagnosis.

Hard part

How to build a product that makes it accessible and intuitive for the elderly and middle-class in India.

Learned

AI can never be democratised if it is not inclusive of language and the communities that dialogue exists in. Technology literacy is very low in India, which is why I embarked on a journey of making 3 prototypes, the last being the most convenient for a typical Indian user.

Awards

• CREST Gold Award ↗
• Blue Ocean Entrepreneurship Competition ↗: Top 500 Finalist of 23,000+
• Recognised by India’s State Government

An AI judge that’s allowed to say “it depends”.

Built

A 5-stage pipeline on top of Stanford’s Co-STORM model. It contains fixed rules to pick a verdict, and is not a standalone model.

Hard part

My test cases and actual user inputs had a difference in accuracy rates. To fix this, I found a filter that was faulty which threw away a quarter of the conflicts, so I deleted it.

Learned

In general, AI is great at facts but horrible with judgement. On complex real-world questions, it’s unable to match any perspective that a sentient human brings.

Worked on live Voice AI models for Indian languages. I shipped two integrated agents for them.

Simulated demo · live in beta

● Connected

Save

WhatsApp

Built a voice agent that sends links and QR codes mid-call. It is live in beta on their agentic platform.

Voice-to-cart

You can talk to any website to fill your cart up. Increased checkout conversion by 9% versus regular checkout over just 3 weeks.

Benchmarks

Tested models on cost, latency, and accuracy.

Learned

Training, training, training. It’s how AI models get better. Real progress actually came from improving the model, not from making added features or integrations.

Conducted a field study for MediTrack on whether people would like to hear their medical records being explained in their own language.

Method

Surveyed 103 adults and conducted 9 interviews in a rural hospital in Muzaffarnagar.

Findings

99 of the 103 preferred voice, and 81.6% of them struggled with the jargon and language in their reports, often relying on a doctor to clarify their doubts.

Learned

Voice is the best method for communicating with unfamiliar technology. Using text becomes unintuitive when voice options exist.

Findings

A plain LLM flipped its answer on 2 of 10 comparisons when the order of the options were swapped. Mine didn’t.

Results

I made 4 fixes to the product itself, out of which 3 were my own mistakes.