You x Naseeb

Timeline: 2 months
Role: Frontend Engineer → Full Stack Implementation
Status: Live in Production

Project Overview

Built a modern dating platform featuring AI-powered conversational onboarding, replacing traditional forms with an intelligent chat assistant. The system processes user conversations, generates structured data, and automates application processing with PDF generation and email notifications.

You x Naseeb's chat interface

Tech Stack

Frontend Layer

Next.js - React framework for server-side rendering
React - UI component library
Tailwind CSS - Utility-first styling
Shadcn UI - Pre-built component library

CMS & Backend

Payload CMS - Headless CMS with custom collections
Pages Collection - Dynamic content management
Custom PDF Export - Server-side document generation
Background Jobs/Cron - Async task processing

AI Processing

Together.ai - AI inference platform
GPT OSS 120B - Open-source language model
AI-SDK Integration - Streamlined LLM interactions
Two-Step LLM Pipeline - Optimized for reliability

Storage & Assets

AWS S3 - Object storage for user uploads
IAM Configuration - Secure access management
File Uploads - Selfie storage and retrieval

Email Delivery

AWS SES - Production email service
Nodemailer - Email transport layer
VDM Analytics - Deliverability monitoring
Custom Templates - Configurable via CMS

Database

PostgreSQL - Relational database
Railway Hosting - Managed database service

Infrastructure

Railway - Application hosting and deployment
One.com - Domain registration and DNS management

Architecture Flow

1. User Journey

User visits site → Next.js renders pages from Payload CMS

2. Chat Initiation

User starts chat → Together.ai LLM asks questions in JSON format

3. File Upload

User uploads selfie → File stored in AWS S3

4. Chat Completion

Chat completes (~100 messages) → Background job queued in Payload

5. Data Extraction

Job executes → LLM extracts key-value pairs from conversation

6. Data Structuring

Key-value pairs generated → LLM converts to structured JSON dataset

7. Application Creation

JSON dataset processed → Application record created in Payload CMS

8. Notifications

Application saved → AWS SES sends emails to user and admins

9. Admin Review

Admin views application → Can download formatted PDF via custom button

Technical Challenges & Solutions

Challenge 1: File Storage Sync Issues

Problem: Payload CMS stores files locally by default. Files uploaded in local development didn’t appear in production.

Solution: Integrated AWS S3 with IAM user configuration. All file uploads now route to S3, ensuring consistency across environments.

Learning: Spent time understanding AWS IAM permissions and S3 bucket policies.

Challenge 2: LLM Output Reliability

Problem: Initial YAML output format failed consistently. LLMs couldn’t generate valid YAML despite explicit instructions and formatting requirements.

Solution: Switched to JSON output format. Reduced failure rate but still encountered 70% failure when asking LLM to both extract information AND format it simultaneously.

Final Solution: Implemented two-step pipeline:

First call: Extract key-value pairs from conversation
Second call: Convert key-value pairs to JSON

Result: Failure rate dropped to near-zero. LLMs perform better with single-task instructions.

Challenge 3: Background Processing

Problem: Processing ~100 messages with LLM calls takes significant time. Can’t block user interface during processing.

Solution: Implemented background job system using Payload’s built-in cron daemon. User sees thank you page immediately while processing happens asynchronously.

Learning: Spent 8 hours learning Node.js async patterns, daemon processes, and job queue management.

Challenge 4: Email Delivery Infrastructure

Problem: Railway’s Hobby plan blocks outbound SMTP. Managed email services (Resend, Mailgun, Sendgrid) cost $15-20/month.

Solutions Evaluated:

Upgrade to Railway Pro: $20/month
Resend: $20/month
Mailgun: $15-20/month
Sendgrid: $15-20/month

Final Solution: AWS SES

Cost: Pennies per email (fraction of managed services)
Required: DNS verification, production access request
Result: Approved for production in <24 hours
Implementation: Nodemailer transport with custom templates

Challenge 5: PDF Generation

Problem: Client needed formatted PDF reports from structured application data.

Solution: Created custom Payload CMS element that appears before save actions. Clicking “Download PDF” generates human-friendly formatted document from application JSON.

Implementation: 2 hours of formatting and layout refinement.

Key Technical Decisions

Why Two-Step LLM Processing?

Single-step extraction and formatting failed 70% of the time. LLMs handle single-task instructions more reliably than multi-task prompts.

Why JSON Over YAML?

Despite clear instructions and critical formatting tags, LLMs consistently produced invalid YAML. JSON proved significantly more reliable.

Why AWS SES Over Managed Services?

Cost optimization. Managed services cost $15-20/month for features we didn’t need. AWS SES costs pennies with same deliverability.

Why Background Jobs?

Processing 100+ messages with multiple LLM calls takes 30-60 seconds. Background processing maintains smooth UX while handling heavy compute.

Why S3 Integration?

Payload’s local file storage breaks in multi-environment deployments. S3 provides consistent file access across development and production.

Results

✅ Delivered on time within 2-month timeline
✅ Successfully deployed full-stack application with AI integration
✅ Implemented cost-effective infrastructure (avoided $20/month email costs)
✅ Built reliable LLM pipeline with near-zero failure rate
✅ Automated application processing with background jobs
✅ Created admin-friendly PDF export system

Skills Acquired

Backend Development

Background job processing
Cron daemon configuration
Async task management

Cloud Infrastructure

AWS S3 bucket management
AWS IAM user configuration
AWS SES email delivery
DNS record management

AI Integration

LLM prompt engineering
Multi-step AI pipelines
Output format optimization
Error handling for AI responses

DevOps

Multi-environment deployment
Railway platform configuration
Email service integration

Tech Stack Summary

Frontend:     Next.js + React + Tailwind + Shadcn
CMS:          Payload CMS
Database:     PostgreSQL (Railway)
Storage:      AWS S3
Email:        AWS SES + Nodemailer
AI:           Together.ai GPT OSS 120B
Hosting:      Railway
Domain/DNS:   One.com

Conclusion

This project pushed me outside my frontend comfort zone into full-stack territory. I navigated AWS services, implemented AI pipelines, built background job systems, and delivered a production-ready dating platform with conversational onboarding.

The most valuable lesson: complex problems often need simple solutions. When the LLM failed at multi-tasking, breaking it into sequential steps solved the issue. When email services were expensive, going lower-level with AWS SES cut costs by 95%.

Frontend engineers can absolutely ship full-stack projects. It just requires willingness to learn, systematic problem-solving, and not being afraid to dive into unfamiliar territory.