Blog
Technical writing on ML inference, agent AI, system design, and lessons from building at scale.
AI Agents as Universal Task Solvers: When Verification Changes Everything
A new theoretical framework shows how AI agents can become universal problem solvers — but only in domains where solutions can be quickly verified. Here's what that means.
Multi-Agent LLM Systems: What the Research Actually Shows
Synthesizing findings from four recent papers on multi-agent LLM systems. The takeaway: more agents rarely means better results. Distributed systems theory, failure taxonomies, and Amdahl's Law explain why — and point to when multi-agent architectures actually make sense.
Stateful Stream Processing Without the Infrastructure Overhead
Velo is a Python library with a Rust core for processing short-lived stateful streams. It fills the gap between stateless serverless functions and always-on stream engines like Flink — based on the stream functions concept from arXiv:2603.03089.
Diagnosing Attention Sinks in Transformer LLMs
An experiment implementing tools to measure attention sinks and activation spikes in transformer models — based on the ICML 2026 paper by Sun, Canziani, and LeCun. Plus a vLLM integration that pins sink tokens in the KV cache.
Implementing CHLU: From Paper to Prototype in One Night
A hands-on experiment implementing the Causal Hamiltonian Learning Unit from an ICLR 2026 paper — a physics-inspired neural network layer that uses Hamiltonian dynamics for stable long-horizon prediction. Here's what worked, what broke, and what we learned.
Building a Multi-Agent Chatbot with LangGraph and MCP
How I built HiveChat, a multi-LLM agent chatbot using LangGraph's ReAct pattern with 7 Model Context Protocol tools for RAG, code generation, image understanding, and more.
From PDF to Podcast: Building PodForge
How I built a distributed system that converts PDFs into engaging podcast-style audio using LLMs, real-time WebSocket streaming, and a 4-microservice architecture.
Teaching an AI Agent to Write Songs: Inside SongSmith
Building an autonomous AI agent that creates complete songs from text descriptions. Learn how we combined LLMs, neural vocoders, and beat generation to create a conversational music composer.
Building Vaani — A Programming Language Where You Code in Hindi
I built a complete programming language with Hindi keywords, Devanagari variable names, classes, exception handling, and a Sudoku solver. Here's why and how.
Hello World — Why I'm Starting This Blog
After a decade of building ML systems, agent AI, and creative tools at Amazon — it's time to start writing about it.