🧵 LLM Inference Optimization — a short threadLLM inference looks deceptively simple—run a forward pass, generate tokens, repeat—but at scale it quickly turns into a systems problem dominated by memory, scheduling, and latency rather than raw compute. Metrics like Time-to-First-Token (TTFT), int...Jan 14 · 2 min readrajarshiwrites.hashnode.dev
2025 — Ending on a High note!If I’m being honest, 2025 was a full-blown rollercoaster. So much happened that half the time I genuinely forget what I did and then randomly remember something and go like — oh wait, I did that too? bruh. But yeah, it’s ending on a good note. That’s...Dec 25, 2025 · 5 min readrajarshiwrites.hashnode.dev
🛡️ FinPal — The AI We Built to Protect People From Hidden Loan Traps & UPI ScamsNot everyone checks their phone notifications — but scammers definitely do. We kept seeing people around us fooled by: “Send ₹1 to get refund” messages “0% EMI” loans that quietly hide processing fees + penalties Confusing RBI cybersecurity rules ...Dec 7, 2025 · 4 min readrajarshiwrites.hashnode.dev
Adversarial Linear Bandits — The Battle Where You Only See One NumberIf you hang around theory halls or ML research boards (like the CSA department at IISc), you’ll inevitably see “Adversarial Linear Bandit” scribbled somewhere. It sounds intimidating, but the core idea is brutally simple: You choose a vector.The env...Nov 19, 2025 · 3 min readrajarshiwrites.hashnode.dev
Why Retrieval Beats Reinvention in 2025Fine-tuning an LLM sounds powerful — “let’s train it on our data and make it an expert.”But here’s the catch: in real-world products, RAG (Retrieval-Augmented Generation) almost always wins. If your goal is reliable, updatable, and explainable AI, re...Nov 11, 2025 · 2 min readrajarshiwrites.hashnode.dev
