Principal PM · Enterprise AI · Evaluation & trust
Shwetha Devanga
I work on the part of enterprise AI where trust is a requirement rather than a slogan: governance, reliability, explainability, and the evaluation systems that make a product’s claims falsifiable.
Principal Product Manager — most recently on Veracode’s AI Security Platform, where I took an AI remediation engine from concept to enterprise deployment and defined the evaluation and agentic-remediation strategy for autonomous security fixes.
10+ years across engineering, product, data, and venture building — Bosch, TE Connectivity, HERE Technologies, MATCHi, an Antler founder cohort in Stockholm, and Veracode. Six years operating in Germany and Sweden; currently based in Bangalore. MBA from Frankfurt School of Finance & Management; Bachelors in Computer Science & Engineering.
Currently exploring product roles at companies building intelligent platforms for regulated environments. When I am not doing product, you will find me reading books, writing, painting, tango dancing, clubbing at a spin class or meditating on life & death.
Currently building
rattl — a revenue floor a credit committee can act on
An EV charging site can run exactly as modelled and still be declined, because nobody will state in writing what the downside looks like. rattl contracts that downside as a number.
The hard part was never forecasting revenue — plenty of tools do that. It is being willing to put a figure in writing that you would have to pay against, and to withdraw it when the evidence turns. That single constraint decides the whole product.
Design-partner stage, and said plainly on the site: sites are scored free of charge to build a calibration record, and no insurance instrument exists yet — writing a real floor needs a licensed carrier and one is not signed.
Visit rattl.in ↗Working tools
Things you can actually use
Not case studies — running artifacts. Each one is a decision I had to make, made usable by anyone facing the same one.
Interactive · agentic evaluation
When should an agent be allowed to act?
A live sandbox for the call every team shipping an autonomous agent makes by feel: at what quality does it earn the right to act without a human in the loop?
Open the tool →Interactive · EV infrastructure
Agstya.charge — where should the next charger go?
A charge-siting intelligence prototype: a real map, an ROI model, a price engine — and an honest note on exactly what’s modelled and what’s fake.
Open the prototype →Interactive · insurance & risk
Hemma — the home log that reprices your risk
A 2024 concept rebuilt: a home-repairs log that turns maintenance history into a live insurance-risk signal — with what would kill it, and the cheapest way to find out.
Open the prototype →Product case studies
Product strategy, worked end to end
Two product-strategy studies — market, users, a falsifiable thesis, and a roadmap. The thinking, worked end to end.
Developer tools · adoption
Growing adoption of a code-analysis tool
Top 3 bets to win Java developers — and why adoption is a context problem, not a coverage problem. Includes a clickable prototype.
Read the case + try it →Fintech · market strategy
A trust-first fraud platform for digital payments
Positioning a mainframe as the fraud-proof backbone for digital payments — market, competition, product. Includes a clickable prototype.
Read the case + try it →Writing
The thinking underneath
The tools are these ideas, hardened. Each essay states a falsifiable claim, then tries to break it.
- Stupid Question → Serious System five essays · invisible systems
- Where Can We Add AI? the eval-suite decision
- The Price of Being Wrong calibration & falsification
- The Cockroach Conspiracy incentives & evidence
- The Anatomy of a Framework thinking tools
- The Architecture of Connection ↗ an 8-part series · Substack
Latest from the newsletter