Author

Bhaskar Paratey

CEO & Founder

Bhaskar founded Partech Systems after three decades of building software that had to work the first time — newsroom systems at Reuters, case-management for government departments, and a long run of enterprise projects since. He started the company because he was tired of watching good technology fail for boring, human reasons. He writes here about where AI actually earns its keep, and where it doesn't.

28 articles. ← Back to the journal

May 28, 2026 · 5 min read
Operations

It was fine in March

AI systems don't fail loudly. They get quietly worse while every dashboard stays green, and you find out from a customer. Here's what to watch instead, and which signal is worth more than all the others.

Keep reading →
April 30, 2026 · 4 min read
EU AI Act

The EU AI Act, a year in: what it actually changed for the people building

A year in, the EU AI Act has stopped being a legal-team topic and become a build problem. The teams treating it as a compliance review at the end are paying for it twice. Here's what it actually forces, and the posture that works if you're not a hyperscaler.

Keep reading →
March 18, 2026 · 5 min read
Agentic AI

Agentic AI in production: the demo lies

The demo where an agent books your travel and files your expenses in one fluid run is real. The system you can run for ten thousand customers without a 3am phone call is a different animal entirely.

Keep reading →
February 14, 2026 · 5 min read
LLMOps

What we actually spend on AI, and where it leaks

The token price on the pricing page is the part of LLM cost that's easy to see and least worth obsessing over. The bill that surprises people is everything around the token — retries, evals, human review, context you stuffed in because retrieval was hard. Treat it as unit economics and the surprises stop.

Keep reading →
January 22, 2026 · 5 min read
Small Models

Small models, big deal

Everyone's watching the frontier-model leaderboard. Meanwhile the most useful work I've shipped this year runs on an 8B model that fits on a phone and never phones home.

Keep reading →
December 11, 2025 · 5 min read
LLMs

The exit you have never tested

Open weights versus closed weights is usually argued as a values question. It isn't. It's a question about switching cost, and almost nobody has measured theirs.

Keep reading →
November 10, 2025 · 5 min read
RAG

RAG done right: it's a search problem, not an AI problem

Most RAG projects underperform for the same reason — the team treated retrieval as an AI problem when it was a search problem all along. The model was never your weak link.

Keep reading →
April 02, 2025 · 5 min read
Evaluation

The only question I ask in an AI demo

There's one question I ask in every AI demo, and it ends about half of them. Not because the answer is bad — because there isn't one, and nobody had noticed.

Keep reading →
November 19, 2024 · 4 min read
Data

The data was always the project

Every stalled AI project I've been called into has had the same root cause, and it was never the model. It was that nobody wanted to fund the boring six weeks at the start.

Keep reading →

Get our best thinking, once a month.

One email a month, from a real person. Essays, practical guides, and the occasional recipe. Unsubscribe with one click.

Prefer a reader? Subscribe by RSS instead.