Instagram content generation flow
Takes a product item, campaign direction, and user requirements through a validator, summarizer, content generator, critique loop, and visualizer.
Assignment Final Project
Agentic Customer Support Flow
A full-stack demo that blends customer enquiries, return assessment, and Instagram content generation into one agentic application.
3
main experiences: enquiries, return flow, and Instagram marketing flow
FastAPI
backend routes with streamed agent output and product lookup
Svelte
interactive frontend with a clean dashboard-style interface
Key Features
What the Instagram flow demonstrates
Create return ticket flow
Looks up the customer and product context, checks the seven return guardrails, and scores the return with policy-backed confidence.
Customer enquiry workflow
Classifies email, text, or voicemail enquiries into seven categories, routes them to the right Snowflake procedure, and drafts a grounded response.
Support operations dashboard
Supports ticket creation, enquiries, return triage, and analytics in the same application.
Snowflake-backed item lookup
Pulls live item metadata from Snowflake so the Instagram workflow can validate the item before generating content.
Agentic iteration loop
Lets the content agent and critique agent work back and forth for a defined number of rounds before the final visual is created.
Instagram Agentic Pipeline
The Instagram agentic campaign pipeline
Mermaid diagram
1. User starts in the frontend
The SvelteKit app lets the user open the Instagram page, look up an item, and enter the campaign details.
2. Frontend sends campaign inputs
The item SK, method section content, and campaign caption are sent to the Instagram workflow endpoint.
3. Validator and summarizer prepare the brief
The validator agent checks item availability and metadata in Snowflake, then the summarizing agent merges those facts with the user requirements.
4. Content, critique, and visual generation
The content agent drafts the campaign, the critique agent improves it in a loop, and the visualizer generates the final image prompt or visual asset.
Why this flow matters
The user only provides a product item and campaign direction. The system then grounds the campaign in catalog data, refines the copy through critique cycles, and finishes with a visual prompt for the final post concept.
Create Return Ticket
The return assessment pipeline
Mermaid diagram
1. User submits a return claim
The agent starts with the customer claim, the item being returned, packaging condition, return quantity, and customer remarks.
2. Snowflake lookup grounds the case
The backend looks up the customer’s recent purchase and item metadata from Snowflake so the assessment starts from real transaction data.
3. Researcher agent checks seven guardrails
The researcher agent evaluates the claim against seven policy checks, asking follow-up questions when the evidence is incomplete.
4. Policy validation adjusts confidence
Each answer is validated against policy logic, and the result is converted into a confidence score that reflects how strong the return case is.
Why this flow matters
This pipeline reduces guesswork by grounding the claim in purchase history, checking policy guardrails, and turning the result into a confidence-driven recommendation for approve, deny, or manual review.
Customer Enquiries
The enquiry classification and reply pipeline
Mermaid diagram
1. User submits an enquiry
The user provides an email, text message, or voicemail transcript describing the problem or question.
2. Voicemail is transcribed when needed
If the input is audio, the system transcribes it to text so the rest of the workflow can treat every enquiry in a consistent format.
3. The enquiry is classified
The response agent identifies one of the seven supported enquiry categories and determines which source-of-truth procedure should be used.
4. Snowflake procedure and draft response
The matching Snowflake procedure returns the validation data, and the agent uses it to generate a grounded draft response for review.
Why this flow matters
This workflow turns unstructured enquiries into a structured support process by classifying the issue, selecting the correct Snowflake-backed procedure, and drafting a response that stays grounded in source-of-truth data.
Documentation
Project docs and source references
Use the links below to jump into the repository documentation and the backend entrypoint that powers the app.