Aman Sharma / Technical product leader
Technical Head of Product at TheAgentic. I sit inside the problem with domain experts in energy, banking, legal, healthcare and logistics, then turn what they know into software that survives production.Note 1 Engineer by training. Forty products taken from research to MVP, fifteen of them live.
Based in Bengaluru, India
The spine
Eight stages in the lifecycle. A filled node means I owned it. A half node means I contributed. An outline node means I advised and someone closer to the domain held the judgment.Note 2 Click a stage to see when, where and what came out of it.
Pick a node. Each one opens a timeline of where that stage actually happened, and filters the work below to match.
Five owned, two contributed, one advised. The states are the honest part.
The arc
I started in backend Python. Authentication, messaging APIs, the unglamorous middle of other people's products. Then machine learning, then platform work on compute orchestration across fourteen clouds, where I learned that most product failures are really scoping failures wearing a costume.Note 3
The move into product was not a move away from engineering. It was a move toward the part of engineering that decides what gets built. I still read the architecture. I still write the PRD myself. I have shipped more than twenty AI and ML products to production across oil and gas, banking, legal, healthcare and logistics, and the pattern holds every time. The expert knows the problem. The engineer knows the machine. Someone has to hold both at once, out loud, in a room where nobody agrees yet.
That is the job I do.
Selected work
Partner names and product names are held back. The domain, the problem class, the decision I made and the tradeoff I took are all here.
Offshore energy
A six agent architecture for continuous monitoring of well integrity, built for enterprise operators and then cut down for mid market ones. The hard call was refusing a single general agent. Failure modes in this domain do not share a diagnostic path, so the split cost us build time and bought us explainability an operator would actually sign off on.
discovery / definition / architecture
Medical devices
A tool that reads a manufacturer's existing documentation against twenty three sections of safety and performance requirements, then reports what evidence exists and what is missing. We built read only connectors and a virtualisation layer instead of copying documents, which killed the data residency objection before it was raised.
discovery / definition / compliance
Retail banking
A white labelled platform for retail lending, with a decisioning engine that runs in advisory mode against a bank's full product catalogue. Every extracted field carries provenance and a confidence signal, because a credit officer who cannot see where a number came from will not use the number.
definition / prototype / architecture
Legal and document heavy work
A reasoning layer for retrieval and inference across large document sets, where the failure mode is not a wrong answer but a confident one. Ranking changed from relevance to cost of being wrong.
architecture / operate
Real estate
A progressive web app that prices and risk scores urban property. I came in on audit, traced defects through the pricing and risk pipelines, and handed back a prioritised fix list rather than a redesign, because the model was right and the plumbing was not.
operate / prototype
Sales enablement
Integration of a reasoning engine into a sales assistant platform, covering preparation before the call and analysis after it. Scope discipline was the whole game here. Two features, well documented, instead of nine.
build / launch
Document intelligence SaaS
A document intelligence framework with a retrieval augmented fine tuning loop running against live customer corpora. Accuracy up eighteen percent, proven by A/B test rather than asserted in a deck.
definition / operate
Cloud infrastructure
On demand and spot orchestration across fourteen clouds and a hundred plus data centres. Customer compute cost down as much as seventy percent. The unglamorous win was a cleanup service for abandoned instances, which is the kind of feature nobody asks for and everybody needed.
architecture / build / operate
Intake
A product manager who only knows their product is a liability in a regulated domain. You need the market, the operator, the regulator, the incentive structure and the reason a person does the irrational thing on a Tuesday. None of that arrives through a standup.
So I read on a schedule, in three buckets. Domain, meaning AI and product and the craft itself. Subject matter, meaning whatever industry I am currently embedded in. And wild, meaning history, physics, cosmology, philosophy, psychology, conflict, art. The third bucket is the one that pays off unpredictably, which is exactly why it stays in the rotation.
The point is not being well read. The point is being able to stand in someone else's position without dragging my own ego into it. Empathy without bias is a research skill, and it is trainable.
What came of it
Side quests
Three things I would still do if nobody were watching: reading, games, and walking around unfamiliar places with a camera.
Books, audiobooks and the few blogs I actually finish. Three buckets, always running.
Shelf still being alphabetisedPlayed, replayed, and happily replayed again. The worlds I keep going back to.
Save file loading, 97%Walking around somewhere new with no plan and a camera. One line per place.
Still in the darkroomWriting
Long and infrequent when there is an argument to make, short and closer to a passing note when the reading is good.
One post a month at most, and only when something is worth passing on. No growth tactics, no drip sequence.
Contact
Best on email. I answer specific questions faster than general ones.