Document pipelines, reconciliation systems, support agents, and natural-language data tools, designed with independent checks, explicit escalation, and per-decision observability. Built to survive production, not just the demo.
Measured in production on client systems I designed and built. Each metric links to the relevant case study. Clients are anonymized under confidentiality agreements.
Real documents, real edge cases, and real compliance rules break prototypes. I build the verification, escalation, and observability layer that lets a system act on its own where it is safe and hand off where it is not.
A system that is confidently wrong 5% of the time gets checked 100% of the time. I design pipelines where the confident path is trustworthy and the uncertain path arrives with evidence, so review effort actually drops.
Two to four weeks to take one capability to a working, instrumented prototype on your data, with a hardening roadmap instead of a sales demo.
Fintech, SME accounting · In production
70 to 90% of transactions automated end to end, 1,000+ per day
An AI-native accounting platform that matches, categorizes, and posts more than 1,000 transactions a day across 10+ businesses, escalating only what it cannot resolve.
Fintech, insurance and annuities · In production
Processing time cut from 4 hours to 8 minutes per application
A four-stage vision pipeline that reads handwritten annuity applications, cutting processing time from 4 hours to 8 minutes, with a monitored optimization loop gated by held-out evaluation.
Best when you have something live or near launch and want it checked before it scales.
Best when you need to prove feasibility or win stakeholder buy-in fast.
Best when you are building an AI team and want senior judgment without a full-time hire.
Every system has an explicit "not confident enough" state that routes to a human instead of presenting a guess as fact.
High-stakes decisions are validated by an independent mechanism, a second model or a rules layer, before they ship or act.
Cost, latency, and accuracy are tracked per decision from the first day in production, not estimated afterwards.
Essay · September 2, 2026 · 10 min read
One model doing the work and an independent mechanism checking it is the single most reliable pattern I know for production AI. Here is how to build it.
Rohit Singh. Eight years in applied AI and machine learning, the last two focused on production agentic systems in regulated financial workflows, owning each system from problem framing through deployment and the reliability work after launch.
Tell me what you are building and where it hurts. I reply within two business days, and if there is a fit we schedule a 30-minute call.