Jhanvi Arora: Production AI · Operational Systems
I turn messy operations into data, intelligence and workflows people actually trust. I start with the business problem, not the model, and own the thing end-to-end, from the first fragmented spreadsheet to the system running in production.
Production systems my team and I built inside a ~2,500–3,000-truck logistics operation across Canada, the US and Mexico. Anonymized, but real.
~70% automated · review team ~20–25 → ~2–3
“Document AI lives or dies on the exception path. Its real value shows up when the rest of the business can ask what it read.”
See how it works→~90% of milestones close themselves
“A stop only becomes a milestone when the evidence agrees.”
See how it works→false alerts ~50% → ~15%
“The model was never the hard part. Making the signals agree was.”
See how it works→one foundation under most of the operational AI
“The highest-leverage AI work is often data architecture. When the foundation reflects the operation, the right AI use cases become obvious, and cheap.”
See how it works→Delivered within a prior employer’s environment. Details are intentionally anonymized. I’m glad to discuss the operating problem, design principles, and production tradeoffs.
How work moves and where it breaks decides the architecture, whether ML earns its place, and how the workflow gets built.
First, make the data trustworthy.
I model the events and flows a business needs to trust its own numbers.
Then, find the patterns that pay.
Forecasting, retrieval, anomaly detection — only where it measurably improves a decision.
Finally, make it a workflow people use.
Rules, confidence handling, exceptions, feedback loops. A production system, not a demo.
Case in point: a yard check-in agent I shipped. In hindsight, most of it should have been plain workflow logic, with AI kept for the genuinely ambiguous cases. I’d build it that way now.
Every role taught me the same thing from a different side: people make judgment calls with too little information, and the model is rarely the hard part. The data and the exception path are.
AI work for operations teams, one workflow at a time.
2026
I sketch when my brain needs to stop being an engineer. I read to understand people as much as problems. And I'm unnecessarily serious about coffee. It all feeds the same instinct: watch closely, find the pattern, make something better.
The reading stack
A novel about finding the bottleneck. Every operation I’ve worked in had one. Usually it wasn’t the model.
The Goal
Eliyahu Goldratt
People run on scripts, so do organisations. Spot the pattern and the workflow designs itself.
Games People Play
Eric Berne
Decisions aren’t right or wrong; they’re bets under uncertainty. So is shipping a model.
Thinking in Bets
Annie Duke
yes, I drew these.
reads the whole board before moving
my favorite model is the one we didn't need
happiest in the exception queue
reviews code daily. kindly, mostly.
steady when the deadline is today
still can't draw hands right
unnecessarily serious about coffee ☕
The coffee test
Start with the problem, add only what earns its place, and don’t over-engineer it. Tap a few ingredients, and I’ll name what you just invented.
The shelf
A blank system. Elegant, useless. Add something.
I work two ways: embedded, as your Lead or Head of AI, or on one workflow at a time, starting with a diagnostic that sometimes ends in ‘don’t.’
Three ways in, depending on how far you want to take it.
Best for
You suspect AI could help somewhere, but don’t yet know where, or whether it’s worth the trouble.
Best for
You already know the workflow. You need it designed and shipped as something people actually rely on.
Best for
You need someone who owns AI outcomes continuously, not a report and a handoff.
Have an operational workflow that is slow, manual, or judgment-heavy?
Let’s find out if AI is actually the right lever, before either of us spends a cent finding out the hard way.
A role to fill, a workflow to fix, or a good argument about whether AI belongs in the room at all. Pick one below.
