RAG Is a Search Problem Before It Is a Generation Problem
Better prompts cannot rescue weak retrieval. A practical framework for relevance, ranking, and evaluation before the model sees a token.
Independent technical journal
Sujithra's
Field Notes
Useful notes on building AI systems that survive contact with reality.
Volume 01 · 15 essaysAI · Software · Data · Interfaces
Latest essays
LoRA Adapters Are a Serving Primitive
Think beyond fine-tuning: adapters change how specialized models can be packaged, routed, versioned, and operated on shared infrastructure.
Zeros Are Data: Forecasting Intermittent Demand
Sparse demand is not missing demand. What long runs of zero reveal—and why standard forecasting intuition breaks.
The archive
Things I wanted to understand properly.
Essays on AI, software, data, and the engineering details that are easy to skip until they matter.
Agent HarnessesReliability
Agents Need Budgets, Not Infinite Loops
A useful agent loop is not “keep trying.” It is a bounded control system with progress checks, explicit stop reasons, and a resumable state.
Model RoutingInference
Model Routers Turn Benchmarks into Runtime Decisions
The best model is a query-level decision. Routers make quality, cost, latency, and risk part of the request path.
Agent EvalsObservability
The Final Answer Is the Least Interesting Agent Trace
A task can pass through a fragile, expensive, or unsafe path. Evaluate tool choice, arguments, order, state changes, and recovery—not only the final text.
Multimodal AIDeveloper Tools
The Screenshot Is Becoming the New Bug Report
A screenshot can show an AI what the user saw. It still cannot explain what the user expected—or what happened one second earlier.
AI ProductsUncertainty
Your AI Assistant Should Be Allowed to Say “I Don’t Know”
We keep rewarding assistants for producing an answer. Sometimes the most useful product behavior is a well-designed refusal to guess.
Small ModelsInference
Small Models Keep Winning the Boring Jobs
The glamorous demo belongs to the biggest model. The repetitive production task often belongs to the cheapest model that reliably fits.
Local-FirstProduct Design
The Cloud Is a Great Place to Sync, Not to Live
An app feels different when your data lives with you and the server helps devices cooperate—not when the server grants temporary access to your work.
Distributed SystemsCaching
Every Cache Is a Tiny Time Machine
Caching makes systems fast by serving the past. The hard question is not whether data is cached, but how old reality is allowed to become.
Context EngineeringLLMs
Your 200,000-Token Prompt Is a Junk Drawer
A larger context window lets you keep more information nearby. It does not organize that information—or make every token equally useful.
Voice AIInteraction Design
Voice AI Has a Politeness Problem
A voice assistant can recognize every word and still be unbearable to talk to. Conversation depends on timing, interruption, and knowing whose turn it is.
Synthetic DataTraining
Synthetic Data Is a Photocopier, Not a Camera
Generated data can multiply examples of a world a model already knows. It cannot quietly replace observing the world itself.
AutomationHuman Control
The Best Automation Leaves a Handle
Automation feels safe when people can see where it is going, interrupt it, correct it, and continue without starting from zero.
About the author
Building intelligent systems, then writing down what the diagrams leave out.
I'm Sujithra Kathiravan, an AI and machine learning engineer interested in applied AI, model infrastructure, developer tools, and the operational details that turn promising demos into dependable systems.