AI-CRM

• 

  Easy way to interact with CRM using NL and unstructured data

• 

  AI CRM system parse inputs into relevant categories and update CRM, avoiding duplicate entries

• 

  CRM dashboard with insightful charts, interact with them using NL

• 

  Snapshot of a deal stages in Kanban format

• 

  Snapshot of a deal stages in list format

• 

  Deal details with next set of actions to take deal to won/lost stage

• 

  Snapshot of a org in list format

Inspiration

Sales updates live in messy places: Telegram threads, emails, meeting notes, and random docs. We kept seeing the same failure mode: the CRM becomes a “second job,” so it stops being trusted. We wanted a workflow where the CRM updates itself from reality, not from manual form-filling.

What it does

AI-CRM turns unstructured text into clean, reviewable CRM updates.

• Capture from pasted text, uploaded files, or URLs
  
• Extract deals, organizations, contacts, service tags, key decisions, and next steps
  
• Verify the proposed updates before anything touches your database
  
• Recall your CRM with natural language (“What did we agree with REP Labs last week?”)
  
• Visualize pipeline health with a stage distribution widget (no AI dependency)

How we built it

• Next.js (App Router) + TypeScript for the product UI
• Prisma + SQLite for a simple, local-first data model
• Gemini for structured extraction using JSON mode, grounded recall, and chat
• A “proposal workflow” that treats AI output as a draft, not truth

In short:
$$\text{Unstructured Input} \rightarrow \text{Structured Proposal} \rightarrow \text{Human Verification} \rightarrow \text{CRM State}$$

Challenges we ran into

• Reliability of extraction: AI output can drift, so we enforced structured responses and strong validation.
• Data integrity: avoiding duplicates and preventing polluted CRM state required careful review/approval flows.
• UI density: making “verify everything” fast on a single screen without hiding key info took iteration.

Accomplishments that we're proud of

• A human-in-the-loop workflow that makes AI safe for production-like CRM updates.
• Multi-modal capture (text/file/URL) that mirrors how operators actually work.
• A dashboard that demonstrates value even without AI: real pipeline distribution from actual deals.

What we learned

• AI is most useful when it’s treated like a copilot, not an autopilot.
• The best UX for AI systems is “suggest → verify → sync,” with clear confidence boundaries.
• Structured outputs (JSON) + validation dramatically reduce downstream complexity.

What's next for AI-CRM

• Stronger deduplication + entity resolution (contacts/orgs across aliases)
• Team workflows: roles, approvals, and audit trails
• More “judge-friendly” analytics: conversion trends, stage aging, and next-step SLAs
• Optional integrations (Gmail/Slack/Telegram) so capture is truly passive

Built With

• api
• charts
• css
• gemini
• lucide
• next.js
• prisma
• react
• recharts
• sqlite
• tailwind
• typescript