Back to project
Case Study

SynexFlow AI - Tattwir Hackathon Case Study

Public case study for the second-place Tattwir hackathon project, focused on AI forecasting, agentic planning, alerts, and IBP architecture.

Overview

SynexFlow is an AI-driven planning prototype built for the Tattwir hackathon, where the team won second place. It connects forecasting, production planning, inventory awareness, external context, and AI recommendations.

Problem

Business planning depends on sales, inventory, logistics, supplier, market, and operations data. When that data is separate, teams react late to demand changes, shortages, delays, and risks.

Context

The public organization has four repos: synexflow-ai for FastAPI forecasting/agents, tatweer-back for Express/MongoDB, tatweer-front for the dashboard, and tatweer-Landing-page for landing visuals.

My Role

Hackathon team member in the second-place team. The case study focuses on the team-built architecture and AI planning work without claiming solo ownership across repositories.

Goals

  • Create an integrated business planning MVP for the Tattwir hackathon
  • Forecast future sales from historical order totals
  • Generate product planning recommendations with product data, web context, and LLM reasoning
  • Classify and respond to operational alerts with an AI-assisted action plan
  • Expose business entities through a REST API for dashboard and AI-service usage
  • Show planning, forecasting, alerts, and external factors through a product-facing dashboard

Technical Decisions

  • Use FastAPI for the AI service so forecasting and agent workflows can move independently from the main product API
  • Use TimeGPT through Nixtla for sales forecast generation from date and total-price history
  • Use LangGraph for multi-step workflows rather than a single prompt, making product planning and alert handling easier to reason about
  • Use DeepSeek through OpenRouter for LLM calls inside the planning and alert workflows
  • Use an Express/TypeScript API with MongoDB/Mongoose for auth, products, clients, orders, suppliers, warehouses, inventory, alerts, actions, integrations, and forecasts
  • Use Next.js, React, TypeScript, Tailwind CSS, and Chart.js for the dashboard and analytics surfaces
  • Use a separate landing-page repo so the hackathon product could be explained publicly while the dashboard and API stayed modular

Architecture

SynexFlow is split into product, data, and AI layers. The Next.js dashboard calls an Express/TypeScript API through Axios. The API owns auth, validation, uploads, and MongoDB records for sales, products, inventory, alerts, actions, integrations, and forecasts. FastAPI exposes forecasting and planning routes. TimeGPT handles forecasts; LangGraph fetches data, adds context, calls an LLM, and returns recommendations.

Architecture Map

Mermaid
flowchart LR
  Visitor["Landing page visitor"] --> Landing["Next.js landing page"]
  BusinessUser["Business user"] --> Dashboard["Next.js dashboard"]
  Landing --> Dashboard
  Dashboard --> Rest["Axios REST client"]
  Rest --> Api["Express TypeScript API"]
  Api --> Auth["JWT auth, validation, uploads"]
  Api --> Modules["IBP modules"]
  Modules --> Records["Products, orders, suppliers, warehouses, inventory, alerts, forecasts"]
  Api <--> Mongo[("MongoDB and Mongoose")]
  Api --> AiApi["FastAPI AI service"]
  AiApi --> Forecast["TimeGPT sales forecasting"]
  AiApi --> ProductGraph["LangGraph product planning workflow"]
  AiApi --> AlertGraph["LangGraph alert action workflow"]
  ProductGraph --> ProductData["Product endpoint data"]
  ProductGraph --> MarketSearch["Web search context"]
  AlertGraph --> ERPData["ERP alert data"]
  AlertGraph --> MarketSearch
  MarketSearch --> LLM["DeepSeek through OpenRouter"]
  ERPData --> LLM
  ProductData --> LLM
  LLM --> Plans["Integrated plans and action recommendations"]
  Forecast --> Insights["Forecast results"]
  Plans --> Dashboard
  Insights --> Dashboard

Case Study Screenshots

4 views
Production planning surface from the landing repo, showing the calendar workflow and shared planning sidebar.
Forecasting dashboard visual: sales trend chart, KPI cards, product popularity, traffic by location, and marketing metrics.
Alerts and predictions surface for stock shortages, shipment delays, unusual increases, and alert history.
External factor integrations for weather, logistics, financial markets, IoT warehouses, competitor prices, and behavior analysis.

Key Features

  • Sales forecasting route for historical data and forecast horizon input
  • Product planning agent workflow with product data fetch, web search context, and integrated plan generation
  • Alert action workflow with ERP data fetch, alert classification, additional context, and final plan creation
  • Backend modules for IBP entities including products, orders, inventory, warehouses, suppliers, alerts, actions, integrations, and forecasts
  • Dashboard screens for production planning, sales forecasting, alerts and predictions, inventory, products, and reports
  • Landing page that communicates AI-driven business planning and execution
  • Second-place hackathon outcome at Tattwir by Micro Club at USTHB

Challenges

  • Making a broad IBP idea understandable within hackathon time
  • Splitting AI, backend, dashboard, and landing repos without losing the product context
  • Designing agent workflows that fetch context before generating plans instead of relying on isolated prompts
  • Connecting AI outputs to operational screens such as forecasts, alerts, production plans, and external factors
  • Keeping the portfolio wording honest by framing it as a team hackathon prototype

Results / Outcomes

  • Won second place at the Tattwir hackathon hosted by Micro Club at USTHB
  • Produced a public four-repo project with AI workflows, backend API, dashboard, and landing page
  • Added an LLM and agentic systems case study
  • Replaced placeholder copy with repo-grounded architecture, visuals, and public links

What I Learned

  • Agentic systems are easier to explain when each node maps to a business step: fetch data, add context, reason, then produce a plan
  • Forecasting and alert workflows become more useful when they are connected to dashboard decisions
  • Hackathon projects need clear architecture diagrams because the product often spans several small repos
  • A winning result is more useful when paired with concrete technical details