Reehan Ahmed

Swiggy just published a great example of pragmatic ML: predicting customer lifetime value *before* a customer places their first order. The challenge: most new users don't order in their first 30 days, while a small group ends up highly valuable — so the model has to separate look-alike users who diverge later. Their fix: a multi-task MLP with 350+ pre-order features (acquisition channel, device signals, geography, category affinity, payment patterns) and shared hidden layers feeding separate heads for Food and Instamart. The interesting twist — adding order count as an auxiliary prediction task alongside lifetime value didn't just help accuracy, it cut the model's parameters by 63% (363K → 135K). Simpler and better. They also ditched standard regression metrics (MAE/MAPE) in favor of decile-based ranking evaluation, since the real use case is ranking users for ad bidding, not predicting exact values. Result: Spearman correlation above 0.75, feeding directly into Google's target ROAS bidding for acquisition. A nice reminder that in applied ML, the right auxiliary task and the right eval metric can matter more than a bigger model. 🔗 Full writeup: https://lnkd.in/eKE5-ENU #MachineLearning #DataScience #MLOps

1

I recently met a Carnatic singer who's building a brilliant AI product for a problem I didn't know existed: She performs regularly. Classical Indian music. Intricate vocal compositions that require live accompaniment. But the issue is: she can't find accompanists. Traditional Carnatic performances need: → A violinist following her vocal improvisations in real-time → Percussion instruments matching her rhythm patterns → Musicians who understand the raga structure she's working within That's 2-3 trained musicians per performance. But there aren't enough trained accompanists to go around. Especially for: → Practice sessions where she's trying new compositions → Smaller performances that can't afford a full ensemble → Late-night rehearsals when musicians aren't available So she either performs without accompaniment (loses the full sound) or cancels. Her husband (an AI engineer) had an idea: What if an AI could listen to her sing and generate real-time accompaniment? Think of it like having a violinist and tabla player who: → Never miss rehearsal → Instantly adapt to her creative choices → Cost a fraction of hiring live musicians They're working on this product now. This is a pattern I see in niche creative markets: If you have deep expertise in a creative domain most people don't understand - that's an edge. You see problems the rest of us can't even articulate. That specificity is exactly what makes it valuable.

9

A mystery AI model is making serious noise in developer circles. No official announcement. No polished launch video. No roadmap. Yet… benchmarks are surfacing. Early testers are whispering. And the question everyone’s asking: 👉 Is this DeepSeek’s next blockbuster? What’s interesting isn’t just who built it. It’s what it signals. We’re entering a phase where: • Breakthrough models drop without warning • Open ecosystems are outpacing closed ones • Performance is no longer the only differentiator • Cost + accessibility are becoming the real disruptors If this is indeed from DeepSeek, it reinforces a bigger shift: 💡 AI leadership is no longer concentrated 💡 Innovation cycles are getting brutally short 💡 The gap between “research” and “production” is collapsing For enterprises, this changes the game: It’s no longer about betting on one model. It’s about building an architecture that can adapt to any model. Because the next disruption? It won’t be announced in advance. Curious to hear your take: Are we heading toward an “open model dominance” era… or will proprietary giants still control the narrative? #AI #GenAI #DeepSeek #LLM #ArtificialIntelligence #EnterpriseAI #Innovation

3

Reaching out to my LinkedIn verse people for clarifying my doubts and possible clarity to a niggling question.. Perhaps Hamel Husain / Shreya Shankar / Aman Khan / Paweł Huryn whom I follow for all things Evals can help shed light. For a multi-AI agent application using LLM's, how do you prepare the golden data set for Evaluations? Do you have a golden data set for each AI agent in the application or do you treat the AI agent as a block with input and output and have one golden data set for the entire multi-AI agent application. Is Perplexity right? "The right move is usually both: maintain an end-to-end golden dataset for the whole multi-agent application, and add targeted per-agent (or per-decision) golden datasets where you need debugging power and regression safety. This matches the “layered” approach used in agentic evaluation: end-to-end tests validate user-visible behavior, while component/step-level tests isolate where failures occur." #AIEvaluation #AgenticAI #AIEngineering #LLMEvaluation #AIAgents #MultiAgentSystems #GoldenDataset

14

11 Comments

At a certain scale, companies outgrow founder centric execution. What they need instead is: –> institutional leadership –> clearer governance –> and separation between operations and exploration Deepinder Goyal's transition at Eternal fits this pattern. The legacy is already visible: –> a category created –>a platform scaled –> a public company stabilized Stepping back from day to day execution doesn’t reduce impact. In many cases, it changes the altitude at which decisions are made.

9