(FDE) - (Part 1) - Project-1: Explain the Client Requirements

🎯 (FDE) Project-1: Explain the Client Requirements — as a Forward Deployed Engineer 1. INTRO ━━━━━━━━━━━━━━━━━━━━━━ Imagine you have just walked out of a meeting with a new fintech client. What do you tell your engineering team? In this video I play the role of a Forward Deployed Engineer (FDE) and brief my team on a real-world style project: PayV Payments, an RBI-regulated payment gateway handling 4.2 million payments a day, wants to build its OWN fraud-detection LLM. It must score every payment, decide ALLOW / REVIEW / BLOCK, and explain why, all in under 250 milliseconds. I break the client requirement down into seven simple questions: who the client is, what is hurting them, what we will build, why they want their own LLM, the key requirements, the business outcomes, and what "success" looks like. ━━━━━━━━━━━━━━━━━━━━━━ 2. WHAT YOU WILL LEARN ━━━━━━━━━━━━━━━━━━━━━━ ✅ How an FDE turns a long client requirement document into a clear team briefing ✅ Why rule-based fraud systems fail at scale (340 rules, 1 in 19 flags actually fraud) ✅ Why fraud = context + behaviour + timing, not just the amount ✅ How a fraud LLM returns a probability, a decision and a plain-English reason ✅ The two halves of the system: offline training vs. live scoring in 250 ms ✅ Why a company would build its own LLM instead of using OpenAI / other APIs (data privacy, PCI DSS, DPDP Act, cost) ✅ The major requirements: data, features, model, API, speed, security, monitoring ✅ How the client will judge success: 5 test gates + a week-8 go/no-go ━━━━━━━━━━━━━━━━━━━━━━ 3. KEY TAKEAWAYS ━━━━━━━━━━━━━━━━━━━━━━ ✅ Always start with the client's business problem, not the technology ✅ Rising fraud losses AND too many blocked genuine customers: two problems pulling in opposite directions ✅ An LLM must earn its place: it has to beat a simpler XGBoost baseline ✅ Rules don't disappear; they become a safety net after the model ✅ Remember the big three: 250 ms latency · no data leaves the client's cloud · beat XGBoost ✅ Clear targets: 35% less fraud loss, 40% fewer false declines, review queue down from 22,000 to 12,000 a day ✅ The client tests you on data you have never seen, so plan for that from day one 🔗 CONNECT WITH ME ON LINKEDIN: 👉   / saurabh-kamal   ━━━━━━━━━━━━━━━━━━━━━━ #ForwardDeployedEngineer #FDE #LLM #FraudDetection #AIEngineering #MachineLearning #FinTech #GenerativeAI #MLOps #SystemDesign #RequirementAnalysis #Payments #UPI #AI