Shiv Automates

The outbound sales system collapsed. Shiv Automates rebuilds it — one AI agent researches, calls, scores, and follows up with a personalized video. Zero humans. Zero code. Powered by Gemini.

Comment

Inspiration

The outbound sales system has collapsed. SDRs burn out sending 200 cold emails a day. Calls go unanswered. Follow-ups die in inboxes. Entire sales teams exist to do what one well-architected AI pipeline should be doing autonomously — and doing better.

As Chief AI Officer at Ravan.ai, I've deployed AI voice agents for enterprise clients across India and the UAE — real estate, insurance, e-commerce. I kept watching the same pattern: the tech works, but the system around it is still human-heavy. Research is manual. Follow-ups are forgotten. Lead scoring is a guess.

The theme of this hackathon — The System Has Collapsed — isn't a hypothetical for me. It's what I fix for a living. So I built the version of the system I wish existed.

What it does

Shiv Automates is an autonomous AI sales architect. One AI pipeline handles the entire outbound cycle — zero humans, zero code.

Enter a prospect's name, phone, and LinkedIn into the dashboard. The system:

  1. Researches them in real-time — enriches data via PeopleDataLabs, scrapes their LinkedIn posts via Apify, analyzes their company website via Firecrawl, and pulls news/articles via Perplexity Sonar.
  2. Synthesizes all research using Gemini 2.5 Flash into a hyper-personalized briefing.
  3. Calls the prospect via a multilingual AI voice agent (Shanaya — speaks English, Hindi, Hinglish, Punjabi fluently) that opens with a specific callback to their recent work.
  4. Analyzes the call transcript with Gemini — generates a lead score, extracts pain points, maps next steps, updates Airtable automatically.
  5. Follows up on WhatsApp with a beautifully-formatted personalized message, then delivers an AI-generated video of me addressing the prospect by name.

All of this happens in under 3 minutes per lead. A human types one name. The system does the rest.

How we built it

Gemini is the brain across three critical phases:

  • Pre-call context synthesis (turning raw research into an agent briefing)
  • Post-call transcript analysis (lead scoring, pain extraction, next-step mapping)
  • WhatsApp follow-up generation (personalized, on-brand copy in Shiv's voice)

The stack:

  • Frontend: React + Vite + Tailwind, polling Airtable every 2s for real-time UI updates
  • Orchestration: n8n (self-hosted) — every phase is a modular workflow
  • Voice layer: Ravan AI voice agent with dynamic prompt variables
  • Research: PeopleDataLabs, Apify, Firecrawl, Perplexity Sonar
  • Database: Airtable (single source of truth)
  • Follow-up: WhatsApp Business Cloud API + Seedance 2.0 for personalized video
  • Design system: Firecrawl-scraped references from Linear, Attio, Vercel to guide UI

Everything is zero-code in execution — n8n workflows, no backend servers, no custom APIs written.

Challenges we ran into

  • WhatsApp Cloud API's 24-hour window — first outreach requires template messages or a prior inbound. Worked around it for demo flow.
  • Seedance safety filters triggering on multi-reference face prompts — solved by simplifying image references and softening identity language.
  • n8n JSON body escaping when Airtable fields contained quotes/newlines — rebuilt payload construction in Code nodes using JSON.stringify instead of inline templating.
  • AI agent output parsing — model sometimes wrapped JSON in markdown fences. Added a cleaner that strips fences before parsing.
  • Real-time dashboard updates — built a diff-detection polling hook that flashes only newly-populated fields, so users see the pipeline work.

Accomplishments that we're proud of

  • Full 5-phase autonomous pipeline working end-to-end, live.
  • First test call (Prachi Kaushik, Asst. Professor at BPIT): Shanaya spoke Hindi, Punjabi, and English fluently — 2 min 34 sec call, zero human intervention, hyper-personalized opener landed perfectly.
  • Real-time dashboard that visibly updates as fields populate — the "watch it work" moment.
  • Gemini integration is deep (3 phases), not cosmetic.
  • Built solo. No team.

What we learned

  • How to architect agentic pipelines where multiple research signals converge into a single decision layer.
  • How to prompt-engineer Gemini for structured outputs that feed directly into downstream automations (JSON schemas with explicit rules beat freeform generation every time).
  • How to design UI for agentic systems — the hardest part isn't building the AI, it's making the AI's work visible.
  • Temperature tuning matters: 0.2 for structured analysis, 0.6 for message generation.

What's next for Shiv Automates

  • Voice → action loops: the AI agent triggering CRM updates mid-call based on what the prospect says.
  • Multi-channel orchestration: extending beyond WhatsApp into LinkedIn DMs and email.
  • Self-optimizing lead scoring: feeding closed-won/lost outcomes back to retrain the scoring prompt.
  • Productizing this for Ravan.ai clients — real estate agencies and insurance brokers burning on outbound cost.

One AI. Infinite outbound. Zero humans.

Built With

  • airtable
  • apify
  • claude-code
  • firecrawl
  • framer-motion
  • gemini-2.5-flash
  • gemini-api
  • javascript
  • lucide-react
  • n8n
  • node.js
  • openrouter
  • peopledatalabs
  • perplexity-sonar
  • ravan-ai
  • react
  • recharts
  • seedance
  • tailwind-css
  • typescript
  • vite
  • whatsapp-business-cloud-api
Share this project:

Updates