Abhishek K.

I designed and built an AI agent that answers real customers' questions about their own money - live, in production, inside a regulated fintech product. A wrong answer there costs someone real money, so I also had to build a way to prove the answers were right. That second part is what most people skip, and it is the main thing I bring to your project. Most AI projects don't fail at the demo. They fail in month three. The demo still runs, nobody can tell whether it is still correct, and it quietly gets switched off. I work with small and mid-sized companies - the ones where one person is the bottleneck and there is no internal AI team. HOW I WORK 1. Put your company's knowledge where the AI can read it. Right now it is spread across Google Drive, old email threads, a few spreadsheets, and one person's head. I pull it into one place so anyone on your team can ask a question and get the answer your best employee would give. 2. Connect it to the tools you already use - CRM, inbox, helpdesk, billing - so it does the work instead of just talking about it. Drafts the quote. Updates the record. Answers the customer. 3. Prove it is right before anyone relies on it. I write a fixed set of test questions - real ones, plus nasty ones designed to break it - and replay them every time the agent changes. You get a report showing what passed and what failed. AI that is confidently wrong is worse than no AI at all. 4. Train your team and hand it over. No lock-in. If you never speak to me again, it keeps running. WHAT I BUILD - AI agents wired into the systems you actually run on - CRM, support desk, billing, internal ops. Not standalone chatbots. - Document and invoice pipelines - extraction, line-item PO matching, tax and compliance checks, and an exception queue so only the real problems reach a person - Retrieval layers (RAG) over messy, real-world company data - Evaluation harnesses, so you can prove the AI works instead of hoping it does - API and tool integrations, so the agent can take action and not only answer Stack: Python, Node.js, TypeScript, OpenAI API, AWS Bedrock, PostgreSQL, vector databases. THE OPERATING LAYER The agent is rarely the hard part. The hard part is that the answer has to come from somewhere true. So the other half of what I build is the layer underneath: your documents, records and systems turned into something an AI can be trusted to read. Sources stay untouched. Anything derived from them traces back to the file it came from. Rules that matter - a compliance boundary, a tax check, an approval threshold - live in the architecture, not in a prompt anyone can edit. I run my own company on one. On demand it gives me the picture a founder usually keeps in five places: where the money is, what the pipeline looks like, what marketing is shipping, what engineering is stuck on, and what we decided six weeks ago and why. One spreadsheet stays the master, so there is never a second version of the truth. The sync behind it has run 1,708 times since June at 99.2% success, and when a source has nothing new it says so instead of inventing an update. For you that looks like: a briefing that assembles itself, a question you can ask in plain English and get a sourced answer to, and the handful of jobs your team still does by hand every week - done, logged, and checkable. Best fit: document-heavy, rules-bound work where someone can tell whether the output is right. Finance, accounts payable, compliance, advisory, claims, filings. ABOUT ME I am a co-founder and a hands-on builder. Before this I co-founded Zen Delivery in India: we put 55,000+ trucks onto a single logistics network, and we built the invoice-to-PO matching system that Philips India's finance team ran roughly 10,000 supplier invoices a month through. Then we sold the company. So I have run an operation and shipped software an enterprise finance team depended on - and I know the difference between a system that demos well and one that survives Monday morning. HOW I PRICE Fixed scope, fixed price, clear finish line. Most projects start with a paid scoping sprint: I map the workflow, show you exactly what I would build and how I would test it, and you keep that document whether or not you hire me for the build. Tell me the one thing your team wastes the most time on. If I can fix it, I will tell you how. If I can't, I will tell you that too.