hi! i’m Krishna, Krishna Kumar Soni — product manager & builder, Bangalore, India
a product manager with 4+ years in freight and gig economy products handling the unglamorous parts - quotations, auctions, payouts, and the screens people use to earn a living.
these days i’m exploring how AI can help curate, rather than create, what’s worth consuming.
builds
- Kyurationslivecollections worth showing, not just storing. saves from X, Instagram, Chrome, and hands them back to ChatGPT / Claude when needed.
- Beatscriptliveknows when a YouTube video is a song, finds its lyrics & shows them beside it.
- Barmudaliveinfinite user interviewers. one prompt away. conversational AI survey builder.
- Pearmocklivegamified case practice for PMs. mock interviews, scored, rated & ranked.
- BoreScrollXlivehands-free X feed, so you can eat with both hands and keep reading.
- Keyflowlivea Figma plugin to quickly start new projects with the key pages & frames ready in one click, in the design system of your choice.
work
two models in a debate are more accurate than each model alone
As a user, I often copied answers between Claude and ChatGPT. I gave one model's answer to the other and asked it to check. The result often felt better. But there was no way to objectively measure it. So, I built Chorus, a harness that runs different models on the same task. Inside Chorus, I built Consensus. Consensus makes two models deliberate under fixed rules: → Each model answers alone first. → For each answer, they give a confidence score. → Then the models critique each other in turns. → A model changes its answer only when it finds an argument that fails. → GPT-6-Sol and Opus 5.5, both at high effort
credit scores, dogs and traffic behaviour
Dogs don’t become loyal because we hit them. Instead, they learn fear or become aggressive. But reward them for doing right, and they choose you for life! Humans are no different. B.F. Skinner, (known for his research in Behaviorism), in his book Science and Human Behavior notes that punishment suppresses a behavior temporarily but fails to build the desired behavior and instead ends up producing avoidance, fear, and hostility toward the punisher. We know this now for over 7 decades since the above gentleman established it, and yet we fail to implement the finding in our real life. In 2025 alone, 70 Lakh+ traffic violations happened in Bangalore raking in ~Rs. 250 Cr in fines. That’s a 200%+ jump in the fines collected in 2024.
why platforms show ratings but not reviews
Why would a food delivery, grocery, or fresh snack platform show ratings… but avoid showing full reviews? And how it leads them to spend more in marketing. It's a deliberate strategy to maximise current throughput. Even with today’s ML + AI (perfectly capable of filtering noise or moderating comments), full text reviews reveal something platforms find hard to solve: prep quality, product handling, delivery quality, and customer service. These aren’t stable inputs, and exposing them publicly makes the platform look volatile. At a public company scale, that volatility becomes commercially expensive.
discoverability and efficiency in chat-first interfaces
How do you solve for discoverability and efficiency in natural language chat first interfaces? Today, we're witnessing huge adoption of LLMs in the products we use and many big product builders very bullish on chat first interfaces. I am not. I still think there is huge value in stitching a user experience that directs the user to get things done seamlessly and effortlessly. Whether chat first or not is a debate for another time. Today I have another question for you to ponder - Imagine you are using a chat app to book a flight. You might not realise that you can also book rental cars or set reminders to make your future bookings from the same interface. The app has to somehow let you know about these features. Discoverability problem!
fake reviews are turning shoppers away from medicines
In India, over 42% of online shoppers in tier 2, 3 cities are turning away from purchasing medicines and health and wellness products due to the presence of fake reviews! I was recently reading a PwC India report written in 2024, where I found this data. This is a huge number and lots of monetary value lost due to lack of trust and transparency. Have you ever gone to Amazon reviews and seen very long descriptive well formatted reviews? Yep. Those are most likely the fake ones. How do you spot a fake review? 🔸 They usually have overly promotional language 🔸 There is similar wording across multiple reviews, often posted in quick succession.
India’s radiologist shortage is a teleradiology opportunity
In India, there's a severe shortage of radiologists, with only around 10,000 radiologists for a population of 1.4 billion. That's just one radiologist per 140,000 people. Even worse, 75% of these are concentrated in urban areas. As a result, the rural patients often have to travel long distances or wait extended periods for getting their reports analysed by radiologists. Even if just 1% of India's population needs a radiology test each year, that's a staggering 14 million tests! Moreover, if 50% of these tests are in rural areas, with each patient spending an average of ₹1,000 on travel and expenses, that's a whopping ₹7,000 crore ($910 million) in annual rural patient spend ready for access!!
what would a 10-star Rapido ride look like
We all have had a 5 star auto ride. What would you call a 6 star ride? How about a 7 star? 10 star? Here's my attempt at tearing the fabric beyond a five star experience for Rapido. 🔸 Rapido Roam - City tours with Auto Drivers as trained guides. 🔸 Rapido Radio - Captains control in-auto podcast / audio vibes with crowd-voted playlists. 🔸 Rapido Rescue - Emergency Ride Service for late night no signal zones. 🔸 Rapido Royalty - Auto with AC, red carpet, cushions, mock-throne. “Ride like Royalty” 🔸 Rapido Record - DJ console and mic in the auto, lets riders record their rap or podcast. 🔸 Rapido Rave - Rave Party on Auto!
the transaction timing fallacy in subscriptions and memberships
Everyone talks about “frequency,” “engagement,” and “LTV.” No one talks about when the user actually commits. That single moment before or at the transaction is the one that decides whether your business compounds or leaks. Zomato Gold and Swiggy One chose the wrong side of time. They asked for commitment before delivering value. And in the real world, faith doesn’t scale. MakeMyTrip chose the right one: it aligns payment with perceived benefit. MMT users see checkout-linked offers and add-ons unlock exactly when they’re about to pay in the moment of decision. There are only two kinds of subscription logic on earth:
experience
- Delhiveryproduct managerowned procurement on TransportOne TMS (Transport Management System). worked on RFQs, spot auctions, bidding tools for transporters, autonomous agents for awarding bids and handling exceptions in the auction process.
- Freight Tigerproduct managerworked on freight invoicing and e-POD (proof of delivery) in a TMS for enterprise shippers and owned agentic automations for internal marketplace SaaS used for carrier matching.
- Kuberanixfounderran a product studio, shipped 8 client projects across fashion, OOH advertising and real estate & finance.
- Urban Companyassociate product managerworked on partner finance features for UC Partner App including earnings & payouts, personal loans and insurance for gig workers.
- Lessonpalproduct consultantbuilt & owned the tutor onboarding and payout flows for a marketplace for online tutoring. in college at the time.
guest book
nothing on the wall yet. go first.