Kartik Vij

I turn messy operations into systems that pay for themselves.

Applied AI, automation and product, from the first conversation with a team to the number on the dashboard. Seven systems in production this year.

What I build

I walk into a business with no AI roadmap, find where hours and money are leaking, design the smallest system that stops it, ship it with production discipline, and build the measurement layer that proves it worked. For the past year I have done that from zero at a housing finance NBFC. Seven systems are in production. Each has a number.

The function, in numbers

Property title research, automatedThe legal team's thirteen-year ownership search across state registry portals, run by a system that knows every portal step by step.Call analytics for collections and supportEvery call transcribed, scored and classified, so team leads coach from evidence, risky loans surface early, and conduct problems are caught the same day.Cloud cost intelligenceAn eight-figure annual infrastructure bill that nobody could attribute or forecast, turned into a savings engine with an owner for every resource.KYC document checks, on the deviceBlur detection, OCR, masking, watermarking, liveness and face match moved from paid vendor APIs to the customer's own phone, with face match on the company's own servers.Document intelligence for lending paperworkOne pipeline that learns a new document type, classifies what arrives, extracts the fields operations needs, and says how confident it is about each one.Policy assistant inside the loan origination systemAn assistant that answers sales and credit staff from the right knowledge base, drafts the support ticket when a person is needed, and shows the policy team where training is failing.Underwriting assistant for the personal discussionA field agent interviews the customer; the system transcribes live, fills the hundred-question form section by section, flags the gaps, and runs the eligibility and policy checks.Support dashboard and the knowledge base it producedWhere support requests come from, how long each one really takes and who was waiting on whom, read from the conversations themselves. Six months of that data became the platform's first documentation.Ticket and satisfaction analysis for customer serviceThe customer-service team drops in its ticket and satisfaction exports and gets a month-on-month analysis by channel, category and turnaround, with the findings written up, as a signed desktop application.Access removal for people who leaveWhen someone leaves, the user-access team used to hunt through sheets and pivots to find every account they held. Now the leaver notice is parsed and every access is one list, with what is already removed and what still needs a hand.Engineering time, by process, person and estimateBuilt for the technology head, from the raw work logs in the ticketing system, to show where engineering time goes, by business process, team and person, and how estimates compare to what the work actually took.Nexus, an internal gateway for AI capabilitiesOne gateway that puts OCR, speech-to-text, vision and language models behind a single API, with queued jobs, plugin tools and a low-code workflow layer, so a team can publish a small AI function without a deployment.

Each one is written up under Work: the leak, the constraint, the system, the number.

How I work

Find the leak before choosing the tool

Every system here started as a conversation with the people doing the work, not as a model looking for a use. The leak is measured in hours or money before anything is designed.

One hardened stack for every system

A single reusable base with authentication, roles, security review and monitoring already in place, so each new system ships in weeks and nothing is built twice.

A measurement layer in every build

Each system carries its own view of what it did and for whom. The result is a number the business owner can repeat, not an opinion from engineering.

Open models, tuned for the job

Where an open model, repurposed and tuned for the task, is enough, it runs on our own infrastructure. The bill moves from per-call to per-server and the data stays inside.

The full approach

Before this

Founder twice. Too Decimal built Bhugoal, IoT and AI weather forecasting for farmers: two patents and a national innovation award presented by the Vice President of India. Saaspect, a venture studio, shipped SaaS across pharmacy, construction and legal-tech.

More about me

Writing

Talk to me

If you are building an AI function from zero, or deciding whether to, I would like to compare notes.

Write to me