“I build the operating system for AI inside hyperscale marketplaces.”
At Agoda my remit is Product Owner for Supply GenAI: I decide what the team builds, in what order, and how we’ll know it worked. Before that I spent 8+ years at Amazon, where I built a $200M business from scratch and put GenAI into daily use for 5,000+ operators.
Agoda · ex-Amazon
Hotel supply has a long tail: far more partners than any account team can work with one by one. My job is to decide where AI should help, in what order, and how we’ll measure it. Engineering owns the build.
One intake and one ranked backlog, so the loudest request doesn’t automatically go next. I write the PRDs, keep the Scrum backlog sprint-ready in Jira, and own acceptance criteria and UAT.
By commercial results. The work is still early, so there are no Agoda numbers here. Every metric on this site is from Amazon.
PRDs, MVP scope, outcome definitions and delivery plans.
How an assistant reads commercial context, picks the relevant guidance and drafts recommendations for a person to review.
Communication guardrails, test cases, evaluation criteria and escalation rules.
Reusable agent skills, refined through testing and user feedback.
One backlog across commercial, product, engineering and data, including a self-service workflow for a commercial use case.
I take on commercial problems that don’t have a playbook yet, and leave behind a way of working the organisation keeps using.
At Amazon that meant finding the gap in the data, writing the business case, designing the operating model, getting it shipped across teams, and writing it up as SOPs so the next market could reuse it.
Nearly four of those years were in Munich: first in product analytics for Amazon Pantry EU, then running Pan-EU operations for brands including J&J, Reckitt and P&G. I built and led a 15-person team there, and later managed three direct reports in Bengaluru.
Find the real bottleneck and size it with data.
Write the business case and design the operating model.
Deliver with product, engineering, commercial, data and finance.
Turn it into a playbook the next market can reuse.
All six are from my time at Amazon.
Wrote the business case, hired 15 specialists and onboarded 25 Pan-EU vendors. Scaled to 3 regions and 50+ partners in under 12 months, at a $200M annual run-rate and $1M a year in savings.
Took a GenAI toolkit from proof of concept to production with Engineering and Data Science: 5 regions in 6 months, ~8.6x first-year ROI, 98% adoption. My prompt and LLM-evaluation guidelines became the org’s standard.
A 90-day review of a $70M, 500+ SKU business. Fixed the unit economics on 190 products and grew revenue from $6M to $9.9M a quarter, against ~7% for the category. The playbook then added $12M a year across 10 EU partners.
SOPs, decision rights, readiness certification and weekly-to-quarterly reviews that flagged ~$20M a week of risk before it hit the P&L. KPIs rose 1,000 bps and onboarding went from 8 weeks to 4.
Role-based learning paths and coaching for 8,000+ managers: learner NPS 85, knowledge satisfaction up 1,500+ bps, vendor satisfaction up 1,000+ bps. I also coached one IC to manager in 12 months, the fastest in the org.
Ran $4.3M of Prime Day and Spring Deals campaigns across supply, marketing and partner teams, and cleared 610 affected SKUs in the 5 days before launch to recover $4.1M in revenue.
Two of my LinkedIn recommendations: one from someone I managed, one from a leader I delivered with.
Soham led multiple enterprise-scale programs supporting a large, high-revenue vendor ecosystem, bringing together strategic clarity and strong execution. He designed governance and operating frameworks that reduced partner ramp-up time by 25% while improving satisfaction metrics by 1,000 basis points … He doesn’t just deliver results — he elevates the people and organisations around him.
His ownership, analytical skills and drive to deliver results were instrumental for the success of those projects. He was really active fostering collaboration and teamwork, which boosted team effectiveness and created an excellent work environment.
I built both on my own before joining Agoda, on synthetic data. Neither is deployed anywhere. Each is a single-agent Claude tool-use loop that answers only from its sources, cites them, and escalates when they don’t cover the question.
Prepares a key account manager’s QBR pack for a restaurant chain: the issues against central targets, the plays that fit, each one cited, and a draft opener.
~3–4 hours of prep to ~2 minutes, my estimate on demo data.
Diagnoses a partner’s top commercial issues and drafts the outreach, using only the retrieved SOPs.
The account manager decides what gets sent.
This site is the third. I designed, wrote and coded it with Claude: one HTML file, no framework, no CMS. The Ask Soham assistant is a scripted keyword matcher rather than a language model, so it costs nothing to run and can’t make things up about me.
Hands-on with Claude and Codex in my own prototypes and daily work. At Agoda, engineering builds and owns the production systems.
Bangkok, since August 2026. I write occasionally about making AI useful inside large organisations.
Latest noteLessons from taking a GenAI toolkit to 5,000+ operators at Amazon.
Read the note →If you’re building AI products in marketplaces, or hiring for that kind of work, I’m happy to talk. Email is fastest.