Reco Social Platform
Development & Scaling

Midterm Project Progress Presentation

Ashrith Sathu · 220905192 · Computer Science & Engineering

Internal Guide Dr. Rohini R Rao

Industry Fischer Jordan

Date April 2026

02 / Introduction

Introduction

Fischer Jordan

Strategic engineering consulting firm headquartered in New York. Combines Strategy, Analytics, and Technology for enterprise clients.

My Role

Full-Stack Engineer assigned to Reco Social. Dec 2025 — June 2026. Full ownership: backend services, web apps, admin dashboard.

Reco Social

Mobile-first social media platform merging short-form video with incentivized gaming & shopping. Users earn credits via game installs and in-app purchases.

400K+Users

15KDAU

5Services

03 / Tech Stack

Technology Stack

Backend

Strapi CMS (Node.js), Express 5 + TSOA

Web Platform

Next.js (SSR, API routes, web funnel)

Admin Dashboard

React + TanStack Query + TanStack Router

Mobile App

React Native / Expo (TypeScript)

Database

PostgreSQL + pgvector (HNSW), Redis

AI & Cloud

Google Vertex AI (Gemini), GCP Storage, Sentry

04 / Literature Survey

Literature Survey

Microservice Architecture (Newman, 2021) — Decomposing monoliths into independently deployable services. Feed isolation for highest-traffic endpoints.
Customer Support Automation with LLMs — Rolling memory chain approach for ticket intelligence. Each AI analysis builds on prior context.
Cross-Platform Attribution — AppsFlyer PBA Web SDK bridges web-to-app attribution. GA4 Measurement Protocol for server-side conversion tracking.
pgvector for PostgreSQL — HNSW indexes enable approximate nearest-neighbor search within the same database as content metadata.
Monorepo Tooling (Turborepo) — Incremental builds for multi-service TypeScript projects with pnpm workspace dependency management.

05 / Problem Definition

Problem Definition

1. Feed Latency & Monolith Coupling

Social feed served from monolithic Strapi CMS. ORM-generated queries with excessive JOINs. Feed shared resources with all backend operations.

2. No Structured Customer Support

No ticket tracking, no automated risk assessment, no integration with internal user/transaction data. Zero visibility into support metrics.

3. Administrative Tooling Gaps

No streak management, no IP blocking, no bulk crediting, no Zendesk integration. Required manual database operations for routine tasks.

4. Web-to-App Attribution Gap

No mechanism to track conversions from web ads through to mobile app install and in-app actions. Impossible to measure campaign ROAS.

06 / Objectives

Objectives

Build a standalone Feed Microservice — Decouple social feed from Strapi into independent TypeScript service with pgvector similarity search and Redis caching.
Develop full-stack Zendesk Customer Support — User-facing web form, backend compliance API with AI risk scoring, and admin ticket management dashboard.
Enhance Admin Dashboard — Streak management, IP blocking, mass credits, team access control, payout enhancements, and performance optimizations.
Implement Marketing Attribution — Dual-layer tracking: client-side GTM + server-side GA4 Measurement Protocol, AppsFlyer PBA, Meta Pixel.
Consolidate into Monorepo — Turborepo + pnpm workspaces with git subtree sync for 5 services.

07 / Feed Microservice

Feed Microservice: 6-Stage Pipeline

00Setup

→

01Retrievers

→

02Filters

→

03Ranking

→

04Heuristics

→

05Formatters

Setup: Load user embedding, feature flags, frequency control
Retrievers: pgvector cosine similarity (HNSW, ef_search=2000) + popular candidates (60/40 blend). 3s soft timeout.
Filters: Cross-retriever dedup, seen/sent (Redis), platform filtering

Ranking: Multi-signal scoring — creator boost, BitLabs rank, P/V/Q/S for AMAs
Heuristics: Interleave content types, Fisher-Yates shuffle
Formatters: Deferred hydration, fuzzy product grouping, affiliate links (Amazon/Connexity), GIF generation

5-second timeout race with Redis top-scans fallback. Express 5 + TSOA. Service-to-service auth via X-Internal-Secret.

08 / Zendesk Support

Zendesk Customer Support System

Full-stack across 3 codebases — web form, backend API, admin dashboard

Two-Phase Backend

Phase 1 (sync): Validate payload, enforce 2-hour duplicate buffer, save to DB, return immediately
Phase 2 (async): 8 parallel DB queries for user enrichment, GCP attachment upload, Zendesk REST API sync with 20+ custom fields

AI Risk Scoring

Vertex Gemini evaluates 5-factor SOP: VIP status, prior tickets, IDV, credit amount, install tracking
Risk score 0–100 with action recommendation
Auto-credit if auto_approve — credits awarded + internal note to Zendesk

09 / AI Intelligence

AI Ticket Intelligence

Rolling Memory Chain

Every 3h: fetch recent tickets, load previous report via previous_report_id
Identifies game-specific clusters and platform spikes
Prunes stale issues (>72h unseen)
Severity: Critical (≥15), High (8–14), Medium (3–7)

Cron Schedule

Hourly: Volume report to Slack
3-Hour: Intelligence with memory chain
6-Hour: Anomaly detection vs 7-day baselines

AI Support Chatbot

Vercel AI SDK + Gemini 2.5
Tool-calling architecture with 10 tools
Streaming responses via fetch API
Max 20 turns per conversation

lookup_game · check_purchases
check_bonus_history · check_ticket_history
check_payout_status · check_account_risk
run_sop_evaluation · create_ticket

10 / Admin Dashboard

Admin Dashboard Enhancements

New Features

Streak Management — Config forms, creation dialogs, milestone auto-set
IP Blocking — IPv4/IPv6/CIDR validation, unblock tracking, admin ID logging
Mass Credit — CSV upload, game metadata validation, failure report download
Team Access — Grant/revoke with cache invalidation
Payouts — IDV status, on-hold, descriptive labels

Performance

Lazy loading for heavy components
Virtualized table rows (TanStack Table)
TanStack Query cache tuning
URL-synced filter state

Also Built

Zendesk tickets management UI
SQL Chatbot enhancements
Review leaderboard (gamification)
Content analysis routes

11 / Web Attribution

Web Platform & Marketing Attribution

Dual-Layer Tracking

Client-side: GTM dataLayer for Google Ads conversions. gclid cookie (90-day TTL).
Server-side: GA4 Measurement Protocol with SHA-256 hashed email (Enhanced Conversions)
AppsFlyer PBA: setCustomerUserId links browser to app user
Meta Pixel: Client + Conversions API (server-side)

Web Funnel

FunnelStep state machine: list → redirecting → polling → success
Install polling every 5s for up to 5 minutes
Verisoul session ID for fraud prevention

Web Parity

Post-registration game experience matches native app
Wallet balance, streaks, earnings, how-to-earn
SEO: metadata, deep links, hashtag routing

12 / Monorepo

Turborepo Monorepo Architecture

reco

React Native / Expo (Mobile App)

reco-landing

Next.js (Web Platform)

reco-admin

React + TanStack (Dashboard)

reco-feed-service

Express + TSOA (Feed API)

reco-content-service

Strapi CMS / Node.js (Backend)

All 5 apps added as git subtrees preserving full commit history
Automated subtree catch-up pull workflow for continuous sync
Shared tooling via pnpm workspaces — consistent dependency versions
Shared infra: PostgreSQL + pgvector, Redis, GCP Storage, Sentry + PostHog

13 / Work Done

Work Done So Far

Timeline	Area	Key Deliverables
Dec–Jan	Feed Microservice	6-stage pipeline: pgvector search, deferred hydration, product grouping, affiliate links, GIF generation
Jan–Feb	Admin: Streaks	Streak management, payout enhancements (IDV, on-hold), award dialog with game search
Feb	Zendesk Support	Full-stack: web form + backend compliance API + admin ticket management UI
Feb	Web Analytics	GTM + GA4 server-side, AppsFlyer PBA, Meta Pixel, Verisoul fraud prevention
Feb–Mar	Admin: Ops	IP blocking, mass credit CSV, team access management, performance optimizations
Mar	AI Intelligence	Gemini risk scoring, auto-credit, rolling memory chain, AI chatbot (10 tools)
Mar	Monorepo	Turborepo + pnpm with 5 git subtrees, automated sync workflow
Mar–Apr	Web Parity	Post-reg game parity, SEO, Google Ads server-side tracking

325+Commits

93Backend

92Admin

100+Web

40Feed

14 / Remaining Work

Remaining Work April — June 2026

AI Knowledge Base — Deploy AI engines with knowledge across the entire PostgreSQL database. Enable natural-language querying of platform data for business users.
SQL Query Generation — Build and benchmark an AI-powered SQL generation system for accurate data retrieval from business user questions.
Admin AI Interface — Develop a UI within the admin dashboard for business users to interact with the AI knowledge base, with confidence scoring on query results.
Automated Monitoring — Implement alerting and health dashboards for the newly decoupled microservice architecture. SLA tracking for feed service.

15 / Conclusion

Conclusion

Key Achievements

Decoupled social feed into standalone microservice with pgvector-powered personalization
Built AI-powered customer support with automated risk scoring and auto-crediting
Delivered 6+ admin operational tools reducing manual DB operations
Established dual-layer cross-platform marketing attribution
Consolidated 5 repositories into Turborepo monorepo

Impact

325+commits across 4 repos in 4 months

<1sfeed response time for 15K DAU

3hticket intelligence cycles with memory chain

10AI chatbot tools for automated support

16 / References

References

S. Newman, Building Microservices: Designing Fine-Grained Systems, 2nd ed., O'Reilly Media, 2021.
Zendesk Inc., "Zendesk Developer Documentation: REST API," 2025.
AppsFlyer, "PBA Web SDK Integration Guide," 2025.
Google LLC, "GA4 Measurement Protocol," 2025.
Vercel Inc., "Turborepo Documentation," 2025.
Vercel Inc., "Next.js Documentation," 2025.
Meta Platforms, "React Documentation," 2025.
Strapi Solutions, "Strapi CMS Documentation," 2025.
A. Johnson, "pgvector: Vector Similarity Search for PostgreSQL," 2025.
Vercel Inc., "Vercel AI SDK Documentation," 2025.

Thank You

Questions?

Reco Social PlatformDevelopment & Scaling