Hi, I'm Nandini.

A new contract opportunity arrives as a document of 200 pages or more, setting out what a client needs and asking whether we can deliver it. Someone senior read all of it before we could decide whether to bid: three days each, with our most valuable decision stuck behind our scarcest calendar.

An agent reads the whole document, pulls out what the client genuinely requires, and matches it against what we sell and what we have delivered before. It scores fit out of 10 against the criteria the business already used, then returns a go or no-go with reasoning, our services ranked, three comparable wins, the risks, and a draft reply for approval.

I timed the manual path first, about three days. Then I built an eval set to find where it could go wrong before anyone relied on it, including whether it could tell a real requirement from a service merely mentioned in passing, and designed the fixes the evals called for. After a month loading the project library and three months of iteration, it ran in three to four hours, matching the humans' calls.

A client who needed qualified people on site sent an email. Someone searched for staff with the right credentials, worked out who was free, called each of them, and wrote back. One request became a thread that ran for days, the records sat in several places, and nobody could see the week without asking.

One system for the whole operation, not a scheduling screen. Every member of staff, their qualifications and availability in one record, with expiring credentials flagged and chased automatically. Clients file a structured request, and an AI assistant I built answers their questions or points them to the next step. The team filters by role, date and qualification to see who is cleared and free. Separate tenants, so one client never sees another's work.

I built it in the order that mattered: the data model and response times first, then the internal view, then the client's, and the assistant last, once what sat underneath was solid. 7,500+ work hours now run through it across 50+ client locations, at 92% repeat requests, with a person one click away whenever a client wants one.

A business with no engineering culture and the same manual work repeating across six departments, each convinced its version of it was unique and could not be automated. Building tools for all of them would have solved the quarter and left a maintenance queue pointing permanently at me. The constraint was never capability. It was that nobody there had built software before.

Build first, to earn the right to be believed: 15+ agents and automations shipped into live operations, each one replacing work somebody was actually doing that week. Then 20+ hands-on sessions across the six departments, run against each team's own backlog rather than teaching AI in the abstract, so people left with something of their own already running.

Around 35% of manual effort removed per team, costed from the hours each agent gave back and what those hours are worth: $350K to $500K of capacity a year. Within six months, 65% of the people in those rooms were building their own tools without me, so the programme stopped depending on whether I was in it. That is the number I am proudest of.

A Y Combinator backed AI startup had a product vision and a twelve-week window to put something real in front of users. Nothing in the vision said which parts of it could actually be built, integrated and shipped inside that window, and the date was not moving.

I turned the founders' vision into a buildable scope, reworked the roadmap around what the date could carry, and did the UX, wireframes and mockups myself. Then I held the seam between design, engineering and operations, plus the outside parties behind the enterprise integrations, where most of the schedule risk was actually hiding. The call that mattered was treating the vision as a starting position rather than a specification, and cutting scope twice rather than moving the date.

We launched the MVP inside the twelve weeks at 95% on-time delivery, with the integrations live rather than deferred to a later release. The product went on to reach around 12,000 users. I owned the build and the launch rather than the acquisition, so the number I claim is the date.