Actionable insights and connections curated by @Accel_India to accelerate your startup journey.

Treating your harness as finished is how last year's workarounds end up slowing down this year's model. So, how do technical founders stop their harness from going stale? That is one of the problems The Working Knowledge's debut working paper takes on, authored by @atemyipod This working paper breaks down how to track what each part of your harness is for and test it against every new model. It also breaks down why a good harness should get smaller as the models under it improve. The full working paper drops soon. Get it 24 hours early through → seedtoscale.com/the-working-…
1
34
Your best AI training data may already be inside your product, in the mistakes human experts correct The Working Knowledge’s debut working paper, authored by @atemyipod, shows how to turn production failures into a feedback loop, the data flywheel: instrument the product, use expert-guided error analysis to curate useful cases, turn them into evals, and use those evals to check whether each product change actually improves performance. The full working paper drops soon. Get it 24 hours early through → seedtoscale.com/the-working-…
1
3
43
How much context should an AI product actually give its model? The Working Knowledge’s debut working paper, “How to Build a Vertical AI Product”, authored by @atemyipod lays out a practical framework for making that decision: benchmark where performance starts to degrade, then use retrieval, compression, progressive disclosure or externalized state to manage context. The full working paper drops soon. Get it 24 hours early through → seedtoscale.com/the-working-…
1
4
43
Judging a model by its overall benchmark score is an easy way to make the wrong assumptions about which tasks can actually be automated. So, how do technical founders work out what a model can reliably handle? That is the focus of The Working Knowledge's debut working paper, authored by Tarun Raheja. The playbook provides a practical framework to score your specific tasks against expert answers and apply Harness Engineering to fix the gaps, showing you exactly when to use better prompting, external orchestration tools, or a human in the loop to make jagged models reliable in production. The full working paper drops soon. Get it 24 hours early through → seedtoscale.com/the-working-…
3
25
Assuming a frontier AI model's intelligence scales evenly is the fastest way to break your AI product. So, how do technical founders actually build reliable vertical AI products on top of unpredictable models? That is the focus of “The Working Knowledge” debut working paper, authored by Tarun Raheja. This working paper breaks down Harness Engineering: and the exact prompts, tools, and logic needed to make a jagged AI model reliable in production. It also breaks down how to build a continuous data flywheel and durably retain your moat against the next major model release. The full working paper drops soon. Get it 24 hours early through → seedtoscale.com/the-working-…
3
36
Your strategy can change. Your product can change. Your mission cannot. @viditaatrey on the first principle he'd teach about pivots: a team needs one anchor point that holds still, because pivots take time and doubt compounds faster than results. (1/2)
1
4
170
In 2021, @Meesho_Official walked away from its reseller business model their $300M Series E had been negotiated on. Half the team was disheartened. Investors asked if it could be run as an experiment alongside the old model. It couldn't. Here are five notes from its case study on why the pivot was still the right call: 1. Meesho never sold anything to shoppers. Women running small shops inside WhatsApp groups did. They showed products to their contacts and took the orders and Meesho got paid by the wholesaler. 2. Then phones got cheap to use, and Covid arrived. Shoppers stopped needing a middlewoman. Users who came to Meesho to resell fell from 70-75% to 35-40% between July and December 2020. 3. As Meesho left the model, Flipkart and Amazon walked into it. Flipkart launched Shopsy for small sellers in July 2021. Amazon bought a reselling startup the year after. 4. Most ecommerce companies take ₹10 to ₹30 out of every ₹100 a seller earns. Meesho took nothing. Its sellers were shifting sari falls and elastic bands under ₹200, where there was no margin to cut into. 5. Winning one customer cost the industry ₹1,000. It cost Meesho about ₹200, because resellers had spent years telling people it existed. FY24: ₹45,650 crore in GMV, ₹7,615 crore in revenue, and ₹197 crore in free cash flow, which was a first for a horizontal ecommerce platform in India.
2
5
425
Resellers were grateful. Consumers were not. @viditaatrey on what @Meesho_Official underestimated when it left reselling in 2021: consumers arrive with expectations, and support, quality, and onboarding all had to be rebuilt to earn the trust that resellers gave for free. (1/2)
1
1
160
"This was an economy that existed that both of us had no idea of." @Meesho_Official led the reseller category at ~$1B GMV, with fresh growth capital raised against that exact model. Listen to @viditaatrey on why he shut it down in July 2021 and rebuilt it as a consumer app. Today there is no major reselling app left in India. But the part worth hearing is the two years of doubt between the decision and the proof. Watch Ep. 3 of Legendary Pivots, Out Now: piped.video/Ub9a3hX9Ehc?si=9L5s…
2
14
6,687
Meesho was the leader in reselling when it walked away from the category entirely and became a consumer app, as that was the signal that came from the ground. @viditaatrey, Co-founder & CEO, @Meesho_Official, on the decision and what it took to execute it. Legendary Pivots Ep. 3 trailer, out now.
3
40
40,685
"Operating mentors teach you day-to-day execution and the right frameworks. Investing mentors teach you risk, reward, and capital allocation." @myspinny's Niraj Singh on why founders need both and why he's still learning how to get the mix right. (1/2)
1
1
170
In 2017, @myspinny told an investor it planned to shut down the marketplace model they wanted to fund and switch to a full-stack model. The investor pulled their term sheet. Headcount went from 75 to 15. Niraj Singh mortgaged his house and funded the company himself for two years. Here are five notes from its case study on why the pivot was still the right call 1. Spinny didn't own the cars. It just introduced buyers to sellers. For every ₹100 a car sold for, it kept about ₹2. It needed ₹6 or more to survive. 2. The seller still ran the sale: They picked the viewing time and handed over the keys. Spinny got blamed if it went badly. 3. So Spinny rented a lot and parked the cars itself: Far fewer cars. But it now kept ₹5 out of every ₹100 instead of ₹2. 4. By late 2018, it was selling 100 cars a month without spending a rupee on ads. Buyers were telling their friends. 5. Early on, almost every car came from listings on other websites. Three years later, almost none did. Sellers were coming straight to Spinny. From 15 people and one rented lot to ₹4,657 crore in revenue and a $1.8 billion valuation.
1
1
183
The used car market was crowded and overfunded. @niraj001s didn't treat that as the deciding question. His filter was simpler: must-have or good-to-have? Good-to-have services don't survive the long term. (1/2)
1
2
122