This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

---

BatchTrace Agent — AI-Powered Food Traceability & Incident Intelligence Platform

"Every batch has a story. Find it instantly."

---

What I Built

The Problem

Small and mid-sized food manufacturers, commercial bakeries, packaged-food producers, and craft processors (20–250 employees) face a massive operational nightmare: their supply chain records live in disconnected silos.

• Supplier invoices arrive as scanned PDFs.
• Receiving dock manifests sit in paper binders.
• Batch production travelers and mixing records are logged in Excel sheets.
• Customer dispatches and bills of lading are stored in carrier portals.

When an ingredient supplier issues an urgent quality alert (e.g., "Supplier lot L-204 of Organic Rolled Oats has elevated DON mycotoxins"), quality assurance (QA) teams spend 48 to 72 hours in panic mode digging through spreadsheets to answer three critical questions:

1. Which manufacturing batches consumed this lot?
2. Which finished products and packaging runs contain it?
3. Which customer orders, distribution centers, and grocery retailers received affected shipments?

A delay in answering these questions can lead to catastrophic public recalls, FDA/USDA warning letters, retail fines, and severe brand damage.

The Solution: BatchTrace Agent

BatchTrace Agent is a full-stack, AI-powered industrial intelligence platform that connects disconnected food manufacturing documents into a deterministic traceability graph, reconciles mass balances, and uses an open-weight AI copilot to guide investigations with 100% grounded evidence.

Who Would Pay for It?

• Small to Medium Food & Beverage Producers: Commercial bakeries, snack producers, granola & cereal brands, and dairy processors.
• Co-packers & Contract Manufacturers: Facilities managing dozens of client SKUs requiring rapid audit-readiness.
• Food Safety & SQF/BRCGS Compliance Consultants: Quality managers who need instant mock-recall compliance reports for annual audits.
• Business Model: SaaS pricing tiered from $199/month (Starter: 5,000 batches/year) to $799/month (Enterprise: multi-facility, live automated document ingestion, and compliance audit exports).

---

Demo

• Live Web Application: https://batchtrace-dev-challenge.vercel.app
• GitHub Repository: https://github.com/divyanshigupta786/Batchtrace-dev-challenge-

Live Platform Features You Can Test:

1. TraceSphere 3D Interactive WebGL Network: A procedural 3D supply chain storytelling environment built with Three.js (@react-three/fiber & @react-three/drei) with camera controls and hover inspection.
2. Interactive Live Trace Simulator: Switch between lots (L-204, L-205, L-206) to watch the forward trace pipeline animate in real-time, detecting the synthetic 50 kg mass balance discrepancy.
3. 2D Operational Traceability Explorer (/app/trace): Execute forward traces from ingredient lots and backward traces from customer orders (ORD-101) to primary agricultural suppliers.
4. Side-by-Side Document Review (/app/documents): Raw invoice and receiving logs previewed side-by-side with coordinate-grounded field extraction (page, row, character offsets) and draft/approve/reject actions.
5. Open-Weight AI Investigation Copilot (/app/agent): Query the agent with natural language; inspect live step-by-step tool execution traces (get_ingredient_lot, trace_ingredient_lot, check_quantity_discrepancies, get_source_evidence); view evidence chips; and approve proposed human sign-off actions.
6. Printable Audit Dossier (/app/incidents/INC-2026-001): Audit-ready printable layout (window.print()) with complete provenance citations.

---

Code

divyanshigupta786 / Batchtrace-dev-challenge-

BatchTrace Agent — AI-Powered Food Traceability & Incident Intelligence Platform

"Every batch has a story. Find it instantly."

Explainable, evidence-grounded food supply chain traceability and batch investigation platform designed for food manufacturers, bakeries, and packaged-food producers.

---

1. System Overview & Problem Statement

When an ingredient supplier issues an alert (e.g. "Supplier lot L-204 has elevated DON mycotoxin levels"), quality assurance and plant operations teams typically spend hours or days digging through disconnected spreadsheets, PDF invoices, paper traveler sheets, and carrier bills of lading to answer three critical questions:

1. Which production batches consumed this ingredient lot?
2. Which finished-product lots and packaging runs were generated?
3. Which customer orders, distribution centers, and retailers received affected goods?

BatchTrace Agent solves this through a dual-engine architecture:

• Deterministic Relational Engine: Serves as the immutable single source of truth. Traverses the graph, enforces mass balances, isolates quantity discrepancies, and computes certainty ratings (CONFIRMED…

View on GitHub

Repository Highlights:

• Next.js 16 Full-Stack: App Router, React 19, TypeScript strict mode, Tailwind CSS.
• Trace Engine: Pure deterministic TypeScript engine (src/lib/engine/traceabilityEngine.ts) and standalone Python engine (backend/engine.py).
• Database Schema: 17 normalized relational tables in supabase/migrations/20261017000000_batchtrace_schema.sql with Row Level Security (RLS) policies and B-tree indexes.
• OpenAPI Microservice: Python FastAPI backend with Pydantic v2 schemas and unit tests.
• Automated Test Suite: 5 test suites (22 assertions) running with npm test.

---

How I Built It

1. Open-Source AI Architecture

At the core of the investigation workspace is the Qwen2.5/Qwen3 open-weight model family, run via Ollama for private, on-premise inference and configurable for hosted endpoints.

Rather than treating the LLM as an unconstrained chatbot that might fabricate customer orders or lot codes, BatchTrace employs a strictly controlled tool-calling harness:

• The AI model operates within a finite set of typed, server-side tools:
  • search_records
  • get_ingredient_lot
  • trace_ingredient_lot
  • trace_finished_lot_backward
  • check_quantity_discrepancies
  • get_source_evidence
  • prepare_incident_report
• Zero Hallucination Guarantee: The relational PostgreSQL database and deterministic traversal engine are the single source of truth. The AI model has no direct SQL write access.
• Evidence-Backed Chips: Every factual claim generated in the agent's response is tagged with an evidence reference (record ID, document citation, row number).
• Graceful Deterministic Fallback: If the local Ollama instance or external inference endpoint is offline, the system seamlessly enters Deterministic Fallback Mode, ensuring 100% uptime for plant operators.

2. Dual-Engine Traceability & Mass Balance Reconciliation

• Forward Traversal: Resolves: Supplier → Ingredient Lot (L-204) → Batches (PB-031, PB-032) → Finished Lots (FP-101, FP-102) → Retail Orders (ORD-101 to ORD-107).
• Backward Traversal: Resolves from any downstream customer delivery back to the master production batch, all consumed raw ingredients, and agricultural growers.
• Mass Balance Audit: Automatically compares received inventory against recorded production consumption. For example, Lot L-204 received 1,200 kg, but production logged 800 kg in PB-031 and 450 kg in PB-032 (1,250 kg total), immediately flagging an intentional -50 kg discrepancy.
• Certainty System: Explicitly labels every relationship as CONFIRMED (backed by approved signed traveler sheets), POSSIBLE (suggested by matching rules but pending QA review), or UNKNOWN (missing records).

3. Frontend & Visual Design System

• Curated Industrial Palette: Forest Green (#173D32), Deep Forest (#102C25), Warm Ivory (#F7F6F0), Soft Sage (#DCE8DF), Signal Amber (#E9A23B), and Risk Red (#B94343).
• Motion & Scrolling: Lenis momentum wheel scrolling paired with Anime.js v4 procedural SVG supply chain path drawing and Framer Motion micro-interactions.
• TraceSphere 3D Experience: Restrained procedural WebGL node network using Three.js and @react-three/drei with camera-facing Billboards and SDF Text.

---

Why Does Open Innovation Matter?

Food supply chain and batch execution data represents some of the most confidential commercial intellectual property a manufacturer possesses:

• Proprietary recipe ratios and ingredient formulations.
• Supplier pricing, volume discounts, and vendor identities.
• Yield efficiency metrics, scrap rates, and operating margins.
• Customer delivery volumes, retailer distribution networks, and packaging agreements.

Why Closed APIs Fail This Use Case:

1. Data Sovereignty & Trade Secrets: Mid-sized food manufacturers cannot stream proprietary formulation logs and internal supplier invoices to closed, third-party proprietary API clouds where terms of service permit telemetry logging or training.
2. Air-Gapped & Edge Plant Readiness: Food processing facilities and grain elevators are often located in rural areas with intermittent connectivity. Open-weight models like Qwen running on local hardware via Ollama allow food safety software to run entirely on-premise without internet dependency.
3. Auditability & Explainability: Regulatory bodies (FDA, CFIA, EFSA) do not accept "the cloud AI said so" during a product recall audit. Open-source models paired with deterministic relational constraints provide verifiable, reproducible, and legally defensible recall dossiers.

Open innovation makes it possible to build enterprise-grade intelligence that respects user ownership, guarantees privacy, and delivers reliable decision-support software.

---

My Agent Session

This complete hackathon MVP was designed, planned, developed, tested, and deployed with Google Antigravity:

• Autonomous generation of the 17-table PostgreSQL Supabase schema with Row-Level Security.
• Development of the dual TypeScript and Python deterministic traceability engines.
• Construction of the React 19 / Next.js 16 frontend with Three.js WebGL canvas and Lenis momentum scrolling.
• Implementation of the controlled open-weight model adapter with bounded tool calling.
• Automated end-to-end verification passing all 5 test suites (22 assertions) with npm test and python backend/test_backend.py.

---

My First User

Persona: Elena Rostova, Quality Assurance Manager at Meadow Pantry Products Inc.

We ran an end-to-end simulation using the synthetic dataset for Meadow Pantry Products Inc., a mid-sized organic granola and snack producer.

The Test Scenario:
Elena received a formal notification from Greenfield Ingredients Co. flagging Lot L-204 of Organic Rolled Oats due to an elevated DON mycotoxin advisory.

Elena's Feedback:

"Usually, when an alert like this arrives at 4:30 PM on a Friday, our QA team has to stay until midnight pulling physical batch folders, matching PO numbers in QuickBooks, and calling cold-storage logistics to see where pallets ended up. With BatchTrace, we pasted the lot code, and within 3 seconds it showed both granola batches (PB-031 and PB-032), flagged the 50 kg over-consumption, identified the 5 affected grocery orders (Whole Foods, FreshDirect, Trader Joe's), and drafted an incident report with exact document references. This transforms a multi-day panic into a 5-minute verified response."

---

Prize Categories

• Best Open-Source AI Project (leveraging Qwen open-weight models via Ollama with controlled tool-calling and zero-hallucination deterministic grounding)
• Best Side Hustle / Commercial Viability (high-value, B2B compliance-critical SaaS targeted at small-to-medium food manufacturers)
• Best Developer Experience / Documentation (full OpenAPI specs, 17-table relational migration, complete test suites, and 1-click deployment blueprints)This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle