Ashesh Srivastava

Ashesh Srivastava

I would like to devote my life to working on something that feels like a sport.

Two-time founder — ecommerce operator turned AI tinkerer. I live at the capability overhang: figuring out interesting use cases, discovering what just became possible, learning as I go — and having fun along the way.

~$85K
bootstrapped AI revenue
$593K
peak monthly GMV operated
7
AI products shipped end to end
Email me →Download résumé

The toolkit

A few of many — each one tagged on the ventures and products below, where it was earned.

Product strategyProduct executionCustomer discoveryStakeholder alignmentLLM evalsAgent architecturePrompt engineeringRAG & retrievalRapid prototypingProduction deploymentFull-stack developmentPython & TypeScript0→1 ownership

First venture to now

2018 — 2023
Two consumer brands on AmazonMy first products were physical: A Baby Cherry® (baby essentials) and Decor & More® (home decor) — sourced from factories in China, Vietnam, and India, sold on Amazon, run with a team of eight. Sourcing, listings, pricing, reviews, returns: the whole loop.
Customer discoveryProduct execution0→1 ownership2brands built & trademarked
2022 — 2023
Marketplace automation at scaleA managed-service operation running US Walmart storefronts for client owners, fulfilled from Amazon — 77 client stores on a custom-built platform that automated listing, repricing, inventory sync, and tracking end to end, with a ~35-person ops team hired in about eight weeks.
0→1 ownershipProduction deploymentProduct executionStakeholder alignment$593Kpeak monthly GMV

I sold the business because I saw the rise of AI — and I wanted to make a meaningful contribution to it. Inside ecommerce the signal was unmissable: everything was getting easier except product imagery. Image generation was the bottleneck, and Stable Diffusion had just landed.

2022 — 2025
QuicSnap → Shutter Labs — one problem, solved twicePhotorealistic product imagery that keeps the fine print exact. First as QuicSnap, an AI-plus-humans service — 3D-model the product, generate the scene around it — sold through an outbound machine to clients across the US, EU, and Middle East. Then Flux.1 shipped, the fine-print moat got an expiry date, and the service became Shutter Labs: a self-serve platform (2D→3D capture plus per-product LoRAs), technical cofounders, a Fortune 500 retailer POC — until Nano Banana Pro absorbed the differentiator and we made the trajectory call to stop. The full story, numbers and all, is the flagship case study.
Product strategyCustomer discovery0→1 ownershipProduction deploymentStakeholder alignment~$85Kbootstrapped revenue

Since then I've been building my own products — designed, shipped, and evaluated solo, to stay calibrated on what frontier models can actually do this quarter.

Canvas
The image workspace built on the model that killed my startup — six models, three lanes, the price printed on the button.
Product strategyFull-stack developmentPrompt engineering0→1 ownership
Orbit
Personal feed ranker over 800+ sources I actually follow; everything new, ranked into one 7am digest.
RAG & retrievalLLM evalsPython & TypeScript
blip
News as vertical reels — caption pipeline tuned from 58% to 92% usable output, measured not vibed.
LLM evalsPrompt engineeringRAG & retrieval
JobFairy
Job-search copilot over 31,926 sponsor companies — with a browser-agent bake-off and real cost data.
Agent architectureLLM evalsProduct executionFull-stack development
Astrape
A utility-bill model validated to within $0.17 of the real bill.
Rapid prototypingCustomer discovery

Principles of building

Canvas taught me this the hard way: I started with a huge requirement doc and got lost inside it. Now every build runs through five gates, in order — each one only makes sense after the one before it.

  1. Make the requirement less dumb

    Question every assumption, especially the ones that arrive looking obvious. With agentic coding, this is where 80% of the time should go — building is cheap now; deciding what to build isn't.

  2. Delete the part or the process

    “Just in case” features and tail-end scope get deleted, not deferred. If you never end up adding anything back, you weren't deleting enough.

  3. Simplify and optimize

    Simplifying and optimizing takes the most energy. It feels exciting to save run time or compute time, but if you do it on the wrong workflow, all that effort is wasted.

  4. Accelerate cycle time

    Once the loop is right, run it faster. Keep tightening every pass until things start breaking, then back off one notch.

  5. Automate — last

    Automation freezes whatever it touches, so it comes after everything above has survived. Then take yourself out of the loop — a script, a cron job, an agent. Automate too early and you're just shipping the mistakes on a schedule.

Pixel-art adventurer holding a torch beside a treasure chest in a rune-lit cave