hi, I'm Divya!

I build AI products, and then I go looking for where they break.

I'm a product manager in San Francisco with ten years at Sunrun, GoDaddy, HP, Warner Bros. and Broadcom. On weekends I build my own AI apps. Below are four of them, plus a teardown I shared with Replit's product team.

all real numbers, promise
4AI products built solo
10 yrsbuilding products
2 LLMsin production, with evals
3languages supported
RAG from Scratch answer with citations
the nerdy oneRAG from Scratch
MoodMango check-in
my favourite coloursMoodMango
Ridge Health home screen
the one with evalsRidge Health
Responsible Nagrik home screen
for every colonyResponsible Nagrik
Things I've built

AI Projects

mostly built on weekends, with chai ☕

Each one has the full story: who it's for, what I decided, what broke, and what I'd do next.

live! ✦
Civic tech· responsiblenagrik.com

Responsible Nagrik

Report a pothole in under a minute, then see if anyone's actually fixing your neighbourhood.

Works in English, Hindi and Hinglish. Every locality gets a public status page, so complaints stop disappearing.

Multilingual UXBuilt with Lovable0→1
3languages
16issue types
7regions
4hstatus refresh
Read about it →
Responsible Nagrik home screenResponsible Nagrik issue picker
has evals!
Consumer health· working MVP

Ridge Health

A food and fitness coach for people who've quit tracking apps before.

You speak your meal, plans use the food you actually cook, and a bad week doesn't reset everything. It also has its own admin console for evals and SLOs.

Voice AILLM evalsCompetitive SWOT
2LLMs with fallback
4SLOs tracked
3competitors studied
Read about it →
Ridge Health home screenRidge Health LLM Ops dashboard
v3 · green mango 🥭
Emotional wellness· moodmango.co

MoodMango

A 30-second mood check-in. You name the feeling, it gives something kind back.

This one is mostly about design: ten practices I leaned on, how it changed from v1 to v3, and my own critique of what to fix.

Product designEmotional UXDesign critique
30scheck-in
34named moods
10design practices
Read about it →
MoodMango check-in with reflectionMoodMango mood snapshot
nerd alert 🤓
AI infrastructure· built from scratch

RAG from Scratch

I asked my RAG pipeline "Which plan supports SSO?" It said the free one.

Six visible stages from document to cited answer, two real bugs, and why chunking (not the model) caused most of the wrong answers.

RAGEmbeddingsRetrieval evals
6pipeline stages
113 msend to end
5fixes ranked
Read about it →
RAG retrieval with similarity scoresRAG answer with citations
Products I didn't build, but have opinions on

Case studies

kind, but honest feedback

Teardowns of tools I use a lot, written up as proposals for the teams that build them.

shared with Replit ✉
Developer tools · Product feedback

Replit improvements

Where Replit users burn time and budget, and a three-phase plan to fix it.

I mapped the "token leakage" loop I kept running into while building, then shared a short deck with Replit's product team: guided prompts, smoother handoffs, and cost visibility.

Read the case study →
How I work

The same four steps, every time

I pick the model last. It's rarely the hard part.

1

Talk to people first

Before I write a prompt, I look for the step people skip, fake or give up on.

2

Pick one bet

I write down the one behaviour that has to change, and how I'd know it did.

3

Build the small version

The smallest thing that tests the bet, with evals, fallbacks and cost limits from day one.

4

Measure, then cut

Keep what moves the number. Delete the rest, even the parts I liked.

About

Ten years in product. Still building.

I started out in engineering and customer success, spent a few years deep in data science, and ended up in product. That's why I'm as happy in a SQL query as in a roadmap review.

Outside work, I build my own AI apps. It's how I stay close to the hard parts: real users, models that fail at odd hours, and costs that sneak up on you.

LinkedIn ↗ · Resume on request

Where I've worked
  1. 2024 – now
    SunrunSales enablement and fintech portfolio management for consumer and partner payments · San Francisco
  2. 2022 – 2024
    GoDaddyCommerce, and experiments to drive activation, retention, engagement and monetization · Santa Clara
  3. 2020 – 2022
    HPE-commerce personalization, ROAS optimization and an AI-enabled supply chain platform · Palo Alto
  4. 2019
    Stanley Black & DeckerFreight and shipment cost optimization · Hartford
  5. 2019
    Warner Bros. EntertainmentNatural language processing and customer sentiment analysis · Los Angeles
  6. 2014 – 2018
    BroadcomCustomer success for enterprise software · Hyderabad
What I work with
LLM applicationsRAGAgentic workflowsAI evalsPrompt engineeringVoice AIRecommender systemsNLPExperimentationGeminiGPT-4o-miniLovableReplitPythonSQLAWSGCPSnowflake
Open to Staff, Principal and Group PM roles

Building something with AI? I'd love to hear about it.

I'm always happy to walk through any of these projects, including the parts that broke.

divyachaturvedi189@gmail.com LinkedIn ↗