Shailesh Shivam · Bangalore
Making software that disappears into its own reliability.
In five years I’ve shipped five systems: a FHIR clinical API used by a hundred hospitals, a blockchain indexer watching eight chains for transaction confirmations, an LLM pipeline that finds credit card numbers hiding in log files, a document migration that moved two million files without a single 404, and an algorithm that decides who sits where in a Fortune 500 office.
Different domains, same shape every time — a user, a timeline, a way the thing could fail silently. That repetition is how taste gets built.
The design system here is called Mithila. The motifs are from Madhubani painting: lotuses for purity, peacocks for beauty, fish for prosperity, yantras for structure. Double-line borders, Kachni lattice textures, hard-offset woodblock shadows, vermilion accents. A folk visual language translated into components — a quiet argument that tools can carry an aesthetic of their own.
“This quarter I’m mostly in LLM-reliability-land — teaching systems to notice when they’re guessing.”
At Saltmine the space-planning engine learned to read floor plans in March. This month I’m teaching it to know when its answers are confident vs. merely plausible — which turns out to be where about 30% of the interesting engineering lives. As usual.
Most evenings I’m still on this website (you’re looking at v8, give or take), and a small open-source thing called kala — an engineering skills toolkit for Claude Code. I’ve also been quietly avoiding a long essay about type-driven design that I promised myself in January. Accountability via public rendering.
- Building Occupancy AI at Saltmine — the reliability layer
- Tending This site · kala (Claude Code skills)
- Returning to Kleppmann’s DDIA — third time around, still catching new details
- Avoiding A long essay about type-driven design (sorry, January-me)
Most engineering problems are thinking problems wearing a coding costume. I’ve never regretted the time I spent on a design document. I’ve frequently regretted the time I didn’t.
How I got here
-
Saltmine
2025–PresentSenior Full Stack Engineer
Fortune 500 facilities teams were spending weeks on spreadsheets to figure out where to seat people. I built a stacking algorithm that evaluates thousands of floor plan configurations against weighted constraints in under two seconds. The interesting problem wasn’t the algorithm — it was modeling the constraints so they’d generalize across wildly different companies.
-
CommandK
2023–2024Founding Engineer
First engineer. The problem: sensitive data hides in places you’d never think to look — credit cards in log files, SSNs in test fixtures, emails in debug dumps. I built the detection engine: ML classifiers, LLMs, and regex working together. No single approach was reliable alone. The ensemble was.
-
Mensari
2022–2023Software Engineer
The legacy system polled every wallet every five minutes. I replaced it with event-driven webhooks across eight chains. The design decision I’m proudest of: a chain-agnostic adapter that made completely different blockchains look identical to the rest of the codebase. Adding a new chain went from weeks to days.
-
Clipboard Health
2022Software Engineer
Two million healthcare documents, Cloudinary to S3, zero downtime. The migration itself wasn’t hard — the verification was. I wrote a shadow pipeline that compared both systems under production traffic until the diff rate hit zero for a week straight. Also optimized the GraphQL API (75% faster) through DataLoader patterns and query batching.
-
Cerner (Oracle Health)
2020–2022Software Engineer
My first engineering job. Built FHIR-compliant clinical APIs (50K daily requests) and a document viewer used by physicians in hospitals. This is where I learned that “production” means something different when the user is a cardiologist between patients. It permanently changed how I think about software.
What drives me
Ownership
End-to-end — from understanding the problem to monitoring the solution in production.
Observability
If you can’t see it, you don’t understand it. Logging, tracing, and alerting are prerequisites, not features.
Craft
The right abstraction beats the right implementation. I’ll spend disproportionate time on the data model because it saves everyone grief for the next three years.
Where it started
Amrita School of Engineering
B.Tech in Computer Science & Engineering
2016–2020Four years of foundations — algorithms, operating systems, databases, networks. The kind of CS education that gives you the vocabulary to learn everything that comes after.
By the numbers
Five years, measured.
How I work
Own the whole problem
Take features from requirements through production monitoring
Write before you code
Design documents are where the thinking happens
The right abstraction beats the right implementation
Spend disproportionate time on the data model
If you can’t observe it, you don’t understand it
Observability goes in the design phase
Ship incrementally
Small, reversible changes over big-bang releases