Aashi Thakkar aashi2912

✦ AI Product Manager · Enterprise + Consumer AI ✦

🧠 The short version

I build AI products that work in production — the kind that survive real users, not just a demo. The hard part is the same whether it's enterprise or consumer: evals, guardrails, and knowing what "good" actually means.

My edge: I was a software engineer first - 3 years building real-time analytics platforms and trading tools. That means I don't just write requirements and hope for the best. I understand what's feasible, where the technical risk lives, and how to have real conversations with the teams building the thing.

On the enterprise side I work at the intersection of LLMs, RAG, and agentic workflows, shipping secure, compliant B2B products. On the consumer side I build and ship my own AI - recommenders and rankers like Vibe Route and Blind Spot (below).

📍 Right now

Role    → AI PM · Engineer-turned-PM
Domain  → Enterprise GenAI + Research Tech
         + Alternative Data
Build   → Consumer AI I ship solo
         (recommenders, rankers)
Stack   → LLMs · RAG · AI Agents · NLP
         Embeddings · Vector · Recsys
Focus   → Works in production, past the demo

⚡ Impact - by the numbers

	What changed	How
🚀	60% faster time-to-insight for end users	LLM-powered automation of high-frequency research workflows
⏱️	12 hours → under 2 hours on critical workflows	AI agents handling data-intensive manual tasks end-to-end
📉	60% reduction in manual research effort	Agentic RAG pipelines replacing repetitive analyst work
📈	25% efficiency gain across teams	Research technology enhancements across cross-functional delivery
⚙️	40% platform performance boost	Optimised caching, API architecture, and data-fetching strategies
👥	8,000+ enterprise users served	Multi-agent GenAI platform featured by NVIDIA · showcased at Databricks Data + AI Summit

🛠️ What I work with

AI / ML

Product

Technical

🧪 Things I've built

Consumer AI, shipped end-to-end - where I keep my product judgment honest.

🗺️ Vibe Route

Google Maps tells you the fastest route. This tells you the best one.

AI-powered walking route comparison that scores routes by how they feel — not just how fast they are. Built with LLM scoring, DBSCAN clustering, and NLP sentiment pipelines.

Claude API DBSCAN ML NLP Product Discovery

📺 Blind Spot

80% of what you watch was chosen by an algorithm. This shows you what it's hiding.

Surfaces filter bubbles in Netflix viewing behaviour using cosine similarity, content-based filtering, and the Herfindahl index to measure recommendation diversity.

Claude API Cosine Similarity Herfindahl Index Recsys

💡 How I think about product

"The best AI products don't feel like AI. They feel like someone finally understood the problem."

Ship small, learn fast, kill what's not working — and have the courage to say so.

I believe	I don't believe
🎯 The user's real problem > the stated requirement	🚫 AI is a solution looking for a problem
📐 A good PRD saves 10x the engineering time	🚫 Moving fast means skipping clarity
🔍 Talking to users is never optional	🚫 Internal proxies understand end users
🛡️ Security and compliance aren't afterthoughts	🚫 You can bolt guardrails on later
📊 "Done" means it passed eval, not that it shipped	🚫 Adoption is someone else's problem

I don't believe in building features nobody asked for. For me, AI isn't the product — it's the tool. The product is the problem you're solving, and how seamlessly you solve it. If the user has to think about the AI, you've already lost.

Always up for conversations about AI product strategy, research tech, or building things that actually work.

Navigation Menu