AI Engineer.AI Engineer.

The company's web product needed a clearer interface across the public website, the client app and the admin panel, and AI agent features that clients could use without extra setup.

I redesigned and shipped the UI/UX across those three surfaces, and I develop AI agent features and LLM integrations, including tool-calling through MCP. I also maintain and secure the backend: database migrations, access policies, API fixes and automated checks in CI.

A chat over a company handbook is only useful if people can check the answer. A fluent answer with an invented source is worse than no answer.

Upload an SOP or handbook and ask about it. Chunks split at the document's own headings. Dense search (a multilingual MiniLM model) and BM25 run side by side, are merged with RRF, then a cross-encoder reranks to 8 chunks. The model tags its claims with citations, and the server checks every one against the context that was actually sent, drops the ones it invented, and shows how many were rejected. Each workspace is isolated by a signed cookie and filtered in both search branches. The interface is in English and Indonesian.

Extraction is easy to demo, but a client only trusts it when wrong answers are caught before they reach their books.

A Python pipeline that reads invoice PDFs with free rules first and calls an LLM only when the rules are unsure. Every result is checked: totals must add up, each value must appear in the source text, and ambiguous dates or files with several invoices go to a review queue with the reason. It reads English, German and Indonesian layouts. On a public benchmark of 113 digital PDFs, with the model on every document, 99 were fully correct and the review gate caught all 14 wrong ones. The prompt was tuned on that set, so it is not a held-out result, and scans are not covered yet.

A RAG assistant is easy to demo and hard to trust or afford. I wanted every answer to show which agent produced it and what it cost in tokens.

A Next.js and Supabase pgvector app. A Manager agent answers general questions and hands document questions to a Specialist agent through a search_docs tool call, with no separate classifier. Documents are split along their own headings into 38 chunks, and when nothing clears a similarity floor the Specialist is skipped. Each API call's token usage is stored, and a 42-question golden set checks routing and correctness.

The MBG/SPPG network needed one place to manage billing, payments, approvals, payroll and contracts across its kitchen units.

A Next.js and Supabase dashboard with 15 modules and role-based access. Billing runs itself: a database job creates the invoices every week, a Python service turns them into PDFs, and a WhatsApp gateway sends them out, with reminder and summary emails on a schedule.

Students rely on the academic office (BAAK) for schedules, the calendar and procedures such as taking leave. Schedule answers have to be exact, while procedures are explained across long guides that one table lookup cannot answer.

A FastAPI chatbot with two paths. Regular-expression rules catch schedule, exam and calendar questions and answer them straight from a Supabase database. Everything else goes through RAG: the question is matched against BAAK's guides in Pinecone and an OpenAI model answers from those passages. Deployed to Google Cloud Run with Docker.

The board game needs everyone at one table. I wanted a version friends could play together online from a private room link.

A Vue 3 client and a Fastify + Socket.IO server that owns every game state, so nobody can cheat from the browser. Game rules live in a separate, tested package. Players join from a room code with no account, on phone or desktop, in English or Indonesian.