Skip to content
Coming soonProject in development

(02)AI Engineering · AI Agents · Multimodal AI · Email Automation

An AI agent that understands a support problem, talks to the company by email and follows up until it's resolved.

Role

Founder & engineer

Status

In development

Planned stack

Next.js, React, TypeScript, Python

Coming soon. In planning and early development, so there's no demo or public code yet.

Coming soonPlanned workflow
  1. CustomerProblem + photos
  2. AI agentVision · RAG · draft
  3. ApprovalCustomer says yes
  4. CompanyEmail + follow-ups
  5. ResolvedConfirmed by customer

Human escalation for complex cases

Customer approves the first email

Scheduled follow-ups until resolved

Concept illustration · not a product screenshot

(01)Overview

An AI-powered customer support platform designed to understand customer problems, analyse uploaded images and documents, communicate with companies through email, manage ongoing conversations and coordinate support until the issue is resolved. It is designed for both individuals seeking product support and businesses automating customer service workflows.

(02)Problem statement

Getting a product issue resolved means finding the right contact, explaining the problem with photos and order details, and chasing replies for days. Businesses face the same loop from the other side, answering routine support emails by hand.

Product vision

Describe the problem once, attach a photo or receipt, and let the agent handle the back-and-forth with the company, with you in control of anything that matters.

Use cases

Personal support

  • Report a faulty product with photos
  • Let the agent find the manual and warranty terms
  • Approve the first email, then let it follow up
  • Get notified when the company replies or the issue is resolved

Business support

  • Triage incoming support emails
  • Answer routine questions from product documentation
  • Draft replies for agents to approve
  • Escalate complex cases with full context

Proposed architecture

  1. 01Web appNext.js, React and TypeScript for cases, approvals and conversation history
  2. 02Agent servicePython, FastAPI and LangGraph orchestrating each support case as a stateful workflow
  3. 03Vision and language modelsLocal models through Ollama: Qwen2.5-VL for images and documents, OpenCV for image processing
  4. 04KnowledgePostgreSQL with pgvector for retrieval over product manuals and documents (RAG)
  5. 05EmailGmail API and Microsoft Graph for sending, receiving and threading conversations
  6. 06Jobs and deploymentRedis for scheduled follow-ups and queues, Docker for deployment

Planned agent workflow

  1. Understandthe customer describes the problem and uploads photos or documents
  2. Analysevision models extract product, model and document details
  3. ResearchRAG retrieves the manual, warranty and support information
  4. Draftthe agent writes the first email to the company
  5. Approvethe customer reviews and approves it before anything is sent
  6. Converseroutine replies are handled automatically, with scheduled follow-ups
  7. Escalatecomplex or uncertain cases go to a human
  8. Resolvethe case is tracked until the customer confirms it's resolved

(03)My role

  • Product vision and scope
  • System architecture and agent workflow design
  • Planning the multimodal, RAG and email integrations

(04)Planned capabilities

These are planned capabilities, not finished features.

  • 01AI-powered customer problem understanding
  • 02Image analysis using computer vision and multimodal AI
  • 03Product and document information extraction
  • 04RAG-powered product manuals and knowledge retrieval
  • 05Automated email drafting and communication
  • 06Customer approval before the first external email
  • 07Controlled automatic replies to routine company responses
  • 08Scheduled email follow-ups
  • 09Human escalation for complex issues
  • 10Support case tracking and resolution confirmation
  • 11Personal and business support modes

Planned technology stack

Planned technologies. This list will be confirmed as each part is built.

Next.jsReactTypeScriptPythonFastAPILangGraphOllamaQwen2.5-VLOpenCVPostgreSQLpgvectorRedisGmail APIMicrosoft GraphDocker

Design principles

Never email on the customer's behalf without consent

The first message to any company needs the customer's explicit approval. Only routine replies after that can be automated, and within set limits.

Knowing when to hand over to a human

Refunds, legal questions or anything the agent isn't confident about are escalated instead of answered.

Grounded answers

Product facts should come from retrieved manuals and documents, not the model's memory.

Development roadmap

  1. Phase 1Problem intake, image and document analysisIn progress
  2. Phase 2RAG over product manuals and knowledgeNext
  3. Phase 3Email drafting, approval and sendingNext
  4. Phase 4Automatic replies, follow-ups and escalationLater
  5. Phase 5Case tracking, resolution and business modeLater

Current status

Coming soon

Planning and early development. Architecture and agent workflow are designed; the first phase is being built.

No demo, public code or documentation yet. This page will be updated as phases ship.

Next project

PanditDesk →