Jai PundeJAI PUNDE

Project 1 / 6

I builtWhatsApp AI Sales Agent

becauseleads message at midnight and expect an answer in seconds.

An AI sales agent on WhatsApp that qualifies leads and books appointments straight into GoHighLevel. Rapid-fire messages are buffered in Redis and answered as one, chat memory lives in PostgreSQL and gets compacted with Gemini, and a single CRM tag switches the AI off so a human can take over.

• Understands text, voice notes, images, PDFs and CSVs
• Buffers rapid messages in Redis so it answers once, not five times
• Checks slots, books, reschedules and cancels through GoHighLevel tools
• Replies like a person: read receipts, typing indicators, split bubbles

▸ 5 input types · 7 agent tools · human handoff

Project 2 / 6

I builtAI Call Summarizer

becausenobody has time to re-listen to sales calls.

A GoHighLevel Marketplace app. Call recordings are transcribed with speaker labels, an LLM scores the lead from 1 to 10 with buying signals and red flags, and a formatted note lands on the contact. Calls under a minute are skipped before any money is spent, and each run's token cost is calculated so usage can be billed per call.

• Speaker-labelled transcripts with talk time per speaker
• Strict JSON output: score, objections, next steps, coaching
• Reads real audio length from the file header to skip short calls
• Per-call token costing for usage-based billing

▸ lead score 1–10 · hot / warm / cold

Project 3 / 6

I builtURL to Media SaaS

becausesmall brands can't afford a design team for every ad.

A full-stack SaaS that turns a website URL into branded image and video content. It scrapes and screenshots the site, reads the uploaded brand brochure into a vector store, and has an AI agent write an editable brand document. Every generation is grounded in that brand context through RAG, rendered by image and video models, and published with Blotato.

• AI-written brand document from the website and brochure (RAG)
• Text→image, image→image, text→video and image→video, logo on or off
• Async jobs with callbacks, timeouts and error logging
• Editing the brand doc re-indexes the vector store automatically

▸ 4 generation modes · brand-grounded

Project 4 / 6

I builtAutomated Lead Revival

becauseold leads aren't dead, they're just ignored.

A 24/7 lead revival system with local-presence SMS drip campaigns that win back old leads at around a 20% reply rate. An n8n AI agent sorts every reply into 15 disposition stages and hands interested leads to the AI appointment setter.

• Local-presence SMS drips that run around the clock
• AI sorts every reply into 15 disposition stages
• Interested leads go straight to the AI appointment setter
• Runs as a GoHighLevel Marketplace app with OAuth2

▸ ~20% reply rate · 15 dispositions

Project 5 / 6

I builtStock Report Insights

becauseanalysts shouldn't copy-paste PDFs into emails.

A Python service watches stock news channels on Telegram and extracts clean text from every new PDF report. It hands the text to n8n through a webhook, where an LLM pulls out the key insights and writes personalised update emails for each client.

• Telethon listener on stock news channels
• PyMuPDF text extraction from new PDF reports
• Webhook hand-off from Python to n8n
• LLM-written, personalised client update emails

▸ PDF → insights → inbox

Project 6 / 6

I builtAI Support Bot

becausethe same questions shouldn't need a human every single time.

An AI support bot that answers customers, creates GoHighLevel tickets and resolves about 70% of issues without a human. Everything else is escalated to the team with the conversation and context attached, so nobody starts from zero.

• Answers, triages and opens GoHighLevel tickets
• Resolves about 70% of issues end to end
• Escalates the rest with full context to a human

▸ ~70% resolved without a human