My SciML (MIT CSAIL) Summer Fellowship on ReservoirComputing.jl: continuous-time reservoirs on top of ODEProblems (ContinuousESN, integrator reuse, sparse Jacobians, LSM, and the cold-start bug that started it all).
What reservoir computing is, how echo state networks work, and four model variants in SciML's ReservoirComputing.jl — SVESM, LIFESN, ResESN, RMNResESN.
What I'm learning about GovTech from inside the Build for Bharat Fellowship, and how I got there.
Dual STT pipeline with WebSocket streaming ASR, Silero VAD state machine, IndicConformer optimized 650ms to 107ms, 128 concurrent users with zero errors.
From 7 idle Ollama containers to SGLang with FP8 quantization, EAGLE speculative decoding, and 256 tok/s — inference engineering on a government H200 GPU at CSG Karnataka.
My experience contributing AI features to DIALOG at Centre for Smart Governance, Govt of Karnataka—ISRO inventory, HRMS portals, Form.io, Camunda, and whether low-code is viable in govtech.
How I took an agentic, multi-form Form.io builder from a 60k+ token context monster to a lean, queue-driven system that fits under 16k tokens—by combining RAG, CAG, and RabbitMQ.
AI-assisted coding is no longer interesting because it can write code; it's interesting because it forces you to pick a philosophy of responsibility. The only philosophy that survives contact with real repositories, real maintainers, and real incentives.
So few days back I was presenting an agent based system at an evaluation panel, and the panelists were referring the whole AI agent design as a glorified function call. It got me thinking about the difference between functions and agents. Yeah maybe there is a similarity but the difference is a bit fundamental.
I hear tales of the old days when geeks spent their billable hours rambling through those classic HTML aesthetic dev-forums, stack-overflow and what-not to get their idea into a working code.