Neha Srivastava
A software engineer who likes wrangling code and context windows, building tools, and breaking things.
Currently at Amazon, building tools and infrastructure that help engineering teams ship faster. After hours, I'm usually herding Claude Code through one of my many, many side quests.
Over 15+ years, my work has spanned a lot of different surfaces, including native apps, quality engineering, developer tooling, CI/CD, platform infrastructure, harness engineering, and more recently, the agentic tooling layer that sits on top of all of it. My role is basically to parachute in and out of different projects and teams wherever I can create the most leverage. These days, a lot of that work involves figuring out where AI tooling actually earns its keep.
I've come to enjoy the kind of problems that live at the end of the road, which I've taken to calling the last 10%. It's the gap between code that runs and code that's truly deployable in production, the kind you can trust to keep working on its worst day. That's where I think most of the interesting engineering actually happens. I write about some of it in my newsletter, The Last 10%.
Before landing in health tech at Amazon One Medical, I spent time in travel and consumer search, different kinds of fast-moving environments that taught me a lot about systems people actually rely on.
Fun fact, I started my Masters at USC in Aeronautical Engineering and switched to Computer Science mid-program after falling in love with the way writing code tickles my brain. And I'm so glad I did.
Other things I make
I'm the creator of Flavor Notes, a coffee tracking app for people who take their beans too seriously, and LinoLog, a tool for documenting linocut prints with AI-generated metadata. Apparently I like logging everything, and at this point I've made peace with it being a personality trait.
Outside of work, I'm a printmaker, and I find carving a block one of the few things that actually slows my brain down in the best way. I also care deeply about accessibility and hold this website to that standard.
The Hardest Thing to Teach an Autonomous Agent Is When to Do Nothing
The Solo Hackathon Problem
The Hidden Economics of Flaky Tests
When Chatbots See Dead Contexts
Baby Proofing by Design: What Hostile Architecture Taught Me
Flavor Notes (iOS · Web · CLI)
Coffee tasting journal on web, iOS, and CLI. The flavor wheel drills down across 120+ tags; the CLI takes natural language and works with Claude Code, Cursor, and Copilot.
Multi-turn attribution
Measures which conversation turns actually drove a model's final answer.
Voice eval review queue
A web queue for human review of AI voice evaluation results.