Problem
Operators lose time when troubleshooting depends on scattered PDFs, tribal knowledge, and manual escalation under pressure.
archive case study
Ops-Copilot
An industrial AI assistant for factory operators: retrieval-augmented, document-grounded troubleshooting on Gemini and Genkit, with Firebase-backed workflows and guided operator actions.
Quick technical review
The case study in four points.
Architecture
The product uses Next.js, Firebase Authentication and Firestore, and Genkit workflows on Gemini 2.5 Flash for retrieval-augmented answers over a per-user manual store. Retrieval ranks documents by lexical relevance rather than by embeddings, which…
Read complete note
The product uses Next.js, Firebase Authentication and Firestore, and Genkit workflows on Gemini 2.5 Flash for retrieval-augmented answers over a per-user manual store. Retrieval ranks documents by lexical relevance rather than by embeddings, which the repository states as its main limitation.
Results
Answers operator questions with document-grounded responses.
Verification path
Inspect this project as an operator-systems concept, not as a warehouse analytics repo.
Summaries reuse the project’s documented evidence. Review the complete case study for evaluation boundaries.
Evidence-led visual review
Explore the actual work.
Concept art, dated captures and repository figures are identified separately.
Portfolio illustration
Original portfolio illustration. Diagrams are conceptual; verified methods and results appear in the linked case study. This is not an application screenshot.
Project proof
Verified external evidence
Last verified 2026-10-05
Problem and approach
Problem
Operators lose time when troubleshooting depends on scattered PDFs, tribal knowledge, and manual escalation under pressure.
Architecture
The product uses Next.js, Firebase Authentication and Firestore, and Genkit workflows on Gemini 2.5 Flash for retrieval-augmented answers over a per-user manual store. Retrieval ranks documents by lexical relevance rather than by embeddings, which the repository states as its main limitation.
Workflow logic
The app answers questions from a curated knowledge base, highlights relevant manual sections, and converts procedural outputs into interactive checklists.
Why it matters
The clearest LLM-product evidence in the portfolio: an LLM application with authentication, persistence and guided actions rather than a chat demo (the measured RAG benchmark is in Visual QC). The repository pairs it with a SECOM failure-risk case study, a class-weighted logistic regression at ROC-AUC 0.7659 with 2.89x failure-yield lift over random review. That result is read on a chronological holdout with only 17 failures (PR-AUC 0.1745; the top-10% review budget captures 29.4%), and the 5-fold time-series mean ROC-AUC is 0.6262 plus or minus 0.1146; the repository calls it an offline portfolio case study, not a deployed live model service.
Architecture and analysis
- Packages retrieval-oriented AI into an operator-facing product surface with persistence.
- Keeps the product narrative grounded in factory troubleshooting rather than generic assistant claims.
- Supports the broader portfolio as evidence of applied AI workflow design.
Testing and validation
- Knowledge-base driven answer generation
- Firebase-backed chat persistence and authentication
- Interactive checklist conversion for execution tracking
Operational value
- Answers operator questions with document-grounded responses.
- Highlights the most relevant manual section behind each answer.
- Supports persistent operator history and checklist-style action tracking.
How to inspect
- Inspect this project as an operator-systems concept, not as a warehouse analytics repo.
- Use it to understand the broader AI layer behind the portfolio rather than the lead analytics path.