Built on the platforms enterprises actually run.
Multi-agent systems, running today.
AP Invoice Processing
Extracts, validates, reconciles, and posts invoices end-to-end. Zero manual touchpoints for the routine 80%.
AP Exception Resolution
Resolves the 20-30% of invoices that fail automated matching — fraud detection, tolerance policies, audit trail.
Four things I am most proud of building.
Not every project. The ones that required the most thinking and produced the most durable outcomes.
A progression of thinking, not a timeline of jobs.
Each chapter taught me something the previous one could not.
Learning to think in systems, not scripts
I started as a software engineer writing test automation and RPA bots at banks and a tech startup in India. The work was tactical — automate a form here, validate a file there. But I noticed something early: the bots that lasted were the ones designed around the process, not the UI. That lesson about designing at the right level of abstraction stayed with me through every role since.
Breadth over depth: how many industries share the same problems
Consulting pushed me into domains I would never have self-selected — insurance, manufacturing, logistics, retail — all within a two-year window. Every engagement started with "we are unique" and ended with the same root problems: data silos, manual handoffs, and processes that had never been questioned. That pattern recognition is one of the most transferable things I built during this period.
Where I learned what enterprise scale actually means
A major global tech company brought me in to lead a Finance Transformation program that delivered 60+ automation initiatives across a global team. I moved from building individual automations to building the framework others used to build automations. I founded a Document Intelligence Center of Excellence and mentored 10+ engineers. We documented 20,000+ hours saved annually.
From automation to agentic AI — and what I learned building both
The most important shift in my career happened around 2022 when I moved from orchestrating bots to orchestrating AI agents. I built LangChain-based agentic workflows on n8n, deployed GPT-4 document intelligence across Procure-to-Pay, and designed an AI governance framework to scale responsibly. I also learned where the new failure modes live: an AI agent can return a confident, well-formed answer that is simply wrong — no error thrown, nothing in the logs.
Positions I hold on enterprise AI.
These came from shipping real systems, watching what failed, and figuring out why.
Connect the data first. Rules second. AI on top.
The biggest mistake I see in enterprise AI programs is reaching for the AI layer before the data and rules layers are solid. L1 integration and L2 rules-based automation are not stepping stones you rush through. They are the foundation everything else sits on.
The orchestration layer must be neutral and company-owned.
Every major platform vendor now has a native AI agent — Agentforce, Joule, Cortex Agents. They are good inside their own walls. The layer that reasons across Finance, GTM, HR, and Legal cannot be owned by any one vendor. Build it on open standards like MCP and A2A.
Governance is the product. Not the constraint on the product.
Semantic failure is the new class of production incident most teams have no playbook for. Standard monitoring tells you when a system is down. It does not tell you when an AI agent gave a confident wrong answer that a human acted on.
Measure hours returned to judgment work, not agents deployed.
The metric that matters is: how many hours did we return to people to do the work that requires actual human judgment? That is the proof that AI did something. Everything else is theater.