About

I am Rohit Singh, a production AI architect. I design, build, and operate AI systems for workflows where a wrong answer has financial, operational, or compliance consequences: document processing, accounting automation, support agents, and natural-language access to operational data. The model is not the product in my work. The controlled system around the model is.

Eight years in applied AI and machine learning, including the last two focused on production agentic systems. On every system below I owned problem framing, architecture, hands-on implementation, deployment, evaluation, and the reliability work after launch. Full employment history is onLinkedIn.

How I got here

1. NLP engineering

   Applied language systems: script processing and sentiment pipelines for a media product.

2. Data science consulting

   Predictive systems for customer communication and lead conversion.

3. Personalization and discovery

   Recommendation and semantic search for a travel platform; a 25.6% lift in booking intent.

4. Production agentic AI

   Document intelligence, reconciliation, support automation, and reliability architecture for regulated financial workflows.

Selected results

• 1,000+ transactions a day, 70 to 90% automated end to end.An accounting and reconciliation platform live across more than ten client businesses, at under 2 seconds and roughly half a cent of AI cost per transaction.Case study
• 4 hours to 8 minutes per application.A vision pipeline for handwritten annuity applications, with a monitored optimization loop that promotes prompt changes only after they pass a held-out evaluation.Case study
• 70.2% of support tickets resolved without a human.A multi-agent support system that resolves recurring issues from a structured incident memory and escalates new ones with a ranked, evidence-backed summary.Case study
• 94.7% extraction accuracy, 78.3% less manual review.A dual-model document pipeline where an independent model checks every extraction before it ships.Case study
• 18 verification steps and 700+ automated tests behind every answer.Natural-language access to manufacturing production data that refuses rather than guesses.Case study

Clients are anonymized under confidentiality agreements. Each case study states how its numbers were measured and exactly what my role was.

How I work

• Refuse over guess. Every system has an explicit "not confident enough" state that routes to a human instead of presenting a guess as fact.
• A second check, always. High-stakes AI decisions are validated by an independent mechanism before they ship or act.
• Instrumented from day one. Cost, latency, and accuracy are tracked per decision from the first day in production.
• Compliance is architecture. Data residency, audit trails, and access control are designed in, not bolted on.

I wrote two essays on the first two: Refuse Over Guess and A Second Check, Always.

Working together

I take on a small number of engagements: a one-week architecture review, a two-to-four-week prototype sprint, or an advisory retainer. The fastest way to reach me is the contact form,hello@rohitsingh.ai, orLinkedIn.