Why most AI projects never leave the demo stage
Four gaps between a promising prototype and a system people can rely on.
I ran an experiment to prove myself wrong. The data had other plans !
I always assumed higher temperature equals more hallucination in LLMs. Seems obvious. But I’ve learned that running on assumptions in…
Why most AI projects never leave the demo stage
Four gaps between a promising prototype and a system people can rely on.
What “trustworthy” really means for clinical AI
Useful clinical AI starts with clear boundaries, reviewable evidence, and accountable decisions.
Citations you can check: RAG for research content
Make the evidence behind generated research summaries easy to locate and verify.
Automating 500 product listings without losing quality
An illustrative batch workflow with validation, review queues, and clear publishing boundaries.
Guardrails for clinical summaries that actually hold
Layered checks make missing evidence and summary errors easier to catch before review.
Guardrails for clinical summaries that actually hold
Layered checks make missing evidence and summary errors easier to catch before review.
The eval set you should build before writing a prompt
Start with representative examples, explicit expectations, and a place to record failures.
Multi-agent systems: when one agent is enough
Add separate agents only when the task needs distinct responsibilities and a useful handoff.
Multi-agent systems: when one agent is enough
Add separate agents only when the task needs distinct responsibilities and a useful handoff.
I ran an experiment to prove myself wrong. The data had other plans !
I always assumed higher temperature equals more hallucination in LLMs. Seems obvious. But I’ve learned that running on assumptions in…
Automating 500 product listings without losing quality
An illustrative batch workflow with validation, review queues, and clear publishing boundaries.
Agents for seller operations: what worked and what broke
A sample design review of bounded tools, approval steps, and recoverable seller workflows.
Why most AI projects never leave the demo stage
Four gaps between a promising prototype and a system people can rely on.
The eval set you should build before writing a prompt
Start with representative examples, explicit expectations, and a place to record failures.
What “trustworthy” really means for clinical AI
Useful clinical AI starts with clear boundaries, reviewable evidence, and accountable decisions.
Guardrails for clinical summaries that actually hold
Layered checks make missing evidence and summary errors easier to catch before review.
From notebook to production: a checklist for LLM features
A release checklist for evidence, boundaries, operations, and ownership before a wider rollout.
Citations you can check: RAG for research content
Make the evidence behind generated research summaries easy to locate and verify.
Detecting hallucinated references in scientific content
Separate reference existence, metadata accuracy, and support for the surrounding claim.
New field notes on this topic are on their way.
Written by Keerthi
Senior Data Scientist building production LLM systems. 11 years in tech, based in Chennai.
A newsletter by Keerthi
Beyond
the Demo.
Practical lessons on shipping AI people can trust. Evaluation, RAG, and agents across healthcare, publishing, and ecommerce.