Er. PrashantPhuyal

Fertility treatment is a subject where a chatbot that improvises is worse than no chatbot at all. A question passes an intent classifier, a domain check and a safety check before anything is retrieved, so an out-of-scope question is turned away before it ever reaches the model. Putting that in the prompt would have been a request; putting it in front is a gate.

Answers come from clinic-approved material retrieved out of Qdrant and fed to Gemini with a prompt constrained to IVF. The knowledge base is built from whatever the clinic already had — PDFs, pages on their website, a YouTube channel, spreadsheets — so there's a parser for each rather than a requirement that someone retype it all.

It also takes phone calls. The voice agent runs on LiveKit, and the part that took longest was barge-in: letting the caller interrupt mid-sentence and having the agent stop talking and listen. Without it you have an IVR menu; with it you have something closer to a conversation.

This one is a separate service rather than a feature. Different teams owned the CRM, the booking system and the customer app, and all three needed AI — so instead of building it into any one of them, I built it as its own service the others call over HTTP.

It covers nine areas: the chatbot, voice calls, damage detection from photos, price estimation, provider matching, provider scoring, review sentiment, fraud and fake-review detection, and analytics write-ups. They vary a lot, but anything involving a model goes through one router — none of them import an LLM SDK directly. Gemini is primary and OpenAI picks up automatically when it fails, and because that switch lives in one place, nine features didn't have to care.

The chatbot works in Nepali first, then English and Hindi. The awkward part is that people write Nepali in Latin letters with no agreed spelling, so a word list for detecting it is hopeless — the same word turns up five ways. It looks at word structure instead.

The platform turns interviews, documents and photographs into a book or a biography. The constraint that shapes everything is that it's published under a real person's name, so the AI isn't allowed the final say. It extracts and it drafts; a human checks what's true in between.

Sources are transcribed or read with OCR, then pulled apart into facts. Two people remembering the same event differently is normal in a biography, so the knowledge module looks for contradictions and surfaces them rather than quietly picking one. Only after someone has verified the material does anything get drafted into chapters.

It's also the first thing I've built with a proper evaluation setup — test cases in English and Nepali with metrics attached, so when I change a prompt I can tell whether it got better or just different. I should have been doing that earlier.

Two sides of the same problem: employers write thin job posts because writing a full one is tedious, and candidates get filtered out by applicant tracking systems without ever being told why.

For employers, a few words about a role expand into skills, keywords, responsibilities and a full specification. For candidates, the CV maker writes a summary, suggests skills, scores the CV the way an ATS would and explains what's dragging the score down — then generates interview questions from what's actually in their CV.

None of it is a chat window, so the output has to be valid structured JSON every time rather than prose that happens to parse. That turned out to be most of the work.