Notes on AI engineering, scaling backends, and lessons from building & shipping.
Asked to critique one paper like a collaborator, I landed on an uncomfortable conclusion — the most dangerous harm an AI can do might be the one that feels like help.
My undergraduate thesis journey — a year of kernel tracing, broken signatures, and 700 simulated attacks that taught me the honest answer is usually more interesting than the clean one.
A bug hid in our Django migrations for a year — deletes that ran too early, reads of fields that no longer existed, and circular dependencies between apps. Here's how two simple rules finally killed it.
A system-design deep dive — in the style of 'System Design Interview' — on why money-moving systems are built the way they are, using Formance Ledger as the case study.
Most RAG demos die the moment a real user asks a real question. This is the build log of DocuSense — every blocker I hit and the fix that got me past it: contextual chunking, hybrid retrieval, PDF highlighting, an on-prem LLM, and retrieving images from documents.
A war story from leading backend at one of Bangladesh's fastest-growing EdTech platforms — the attacks, the exhausted servers, and the live-class WebSocket meltdown, and how we fought through each one.
The ILM AI Hackathon problem statement asked for two physics-accurate simulations — projectile motion and the two-body problem. Three days later we'd built both, polished the UI, recorded the demo, and won. Here's how.