AutoCare Platform
AutoCare is a lightweight, offline-first semantic search and quality control portal designed for Maruti Suzuki. It enables factory floor operators, customer care agents, and QA engineers to quickly query standard maintenance manuals, verify engineering drawings, and track model approval pipelines without heavy machine learning dependencies.
⚠️ The Business Problem
Automotive manufacturing and support operations generate massive volumes of technical questions daily (e.g., clearance tolerances, paint processes, bumper replacement specs). Conventional support portals rely on keyword matching, which fails when users query with synonyms or rephrased terminology.
A support agent querying How long is the wiper assembly covered? will fail to retrieve a standard FAQ titled Model 1 Warranty Specifications if there is no direct keyword overlap, increasing agent handle times and quality compliance overheads.
💡 The Semantic Solution
AutoCare processes documentation and user inputs into dense float vector matrices. The system checks meaning instead of exact keywords:
Sentence Embedding Generation
User queries and PDF documents are mapped to 384-dimensional dense vectors using the sentence-transformers/all-MiniLM-L6-v2 ONNX model.
Cosine Similarity Scoring
Similarity matching is calculated in memory using NumPy dot-products, bypassing heavy vector database runtimes and maintaining a RAM footprint of less than 100MB.
Department-Specific Classification
Lexical rules automatically categorize queries into designated departments (Service, Spare Parts, Insurance, Roadside Assistance, or QA) for telemetry mapping.
👤 User Roles & Privileges
Permissions are strictly enforced based on roles configured in the system:
| Role | Department Scope | Available Actions | Permissions |
|---|---|---|---|
| QA Admin | QA & Management | Full CRUD, configuration management, uploader approval, audits. | All (Admin, Approve) |
| QA | Quality Assurance | Upload engineering drawings, query chatbot, review configurations. | View, Upload |
| Engineering | Design & Engineering | Upload specifications (require QA approval), query chatbot. | View, Upload (Pending) |
| Production | Assembly Line / Factory | Read-only chatbot access, check specifications. | View |
Interactive Mock Workspace
Customer Portal Interface: Connects to Maruti support. Select your department, click chips, or type custom queries.
Admin Panel Dashboard: Sign in using the default QA Admin credentials (qa_admin / QAAdmin123!) to inspect live telemetry charts, approve engineering specs, add FAQs, and review audit logs.
KPI Reference manual
Grounded Answer Rate
Definition: The ratio of customer questions answered using certified FAQ entries or approved documents compared to overall queries.
Formula: GAR = (Grounded / Total Queries) * 100
Exclusions: Repeated consecutive exact user inputs or default system checks.
Unresolved Escalations
Definition: Support requests initiated via Thumbs Down reviews or consecutive matching errors that are not set to 'Resolved'.
Formula: Count of logged tickets where status != "Resolved".
Business Rule: Escalation tickets must be resolved within 48 hours to comply with Maruti Quality guidelines.
Detailed KPI Deep-dive
Grounded Answer Rate (GAR)
This is the primary quality health metric monitored by management. A grounded answer means that the customer received an answer retrieved from either the FAQ database (faqs.json) or approved PDF spec manuals (documents.json) with a matching score above 0.45.
Escalation Tickets Breakdown
| Status State | System Behavior | SLA Constraint | Resolution Trigger |
|---|---|---|---|
| New | Ticket logged; customer notification queued. | Unchanged for 12 hours max. | QA Admin assigns or marks In Progress. |
| In Progress | Assigned to investigator. Customer notified. | Total resolution < 48 hours. | Corrective FAQ entry published. |
| Resolved | QA approved resolution. Ticket archived. | — | Ticket state saved. GAR updated. |
Key Platform Features
Semantic Query Matching
Translates natural language synonyms (e.g., "drawings" vs "specifications") to identify matching manuals using MiniLM ONNX vectors.
Ingest quality Checks
Verifies uploads against required structural sections (Intro, Methodology, Conclusions, Safety), enforcing a 60/100 threshold.
RBAC Security Security
Restricts admin dashboard configurations to QA Admins. Engineers are restricted to uploads requiring verification.
Feedback Loop Log
Tracks thumbs ratings, logging user disputes directly into gap queues, triggering immediate QA review alerts.
Customer Chatbot Portal Workflow
Customers select a department and input queries. The chatbot tracks a sliding history window (up to 6 turns). If the user enters a short query (under 8 words), client-side helpers concatenate the previous user turn to preserve context.
Citations UI cards render directly beneath matched document answers, highlighting document names, category tags, page numbers, and exact confidence percentages.
Admin Dashboard Pipeline
The Admin Portal acts as the control hub. Ingested engineering drawings remain in a Draft state until reviewed and approved by a QA Admin. Once approved, the document content is split into page-level chunks, embedded, and dynamically concatenated to the numpy index matrix without requiring a server reboot.
Quality Control Workflows
Engineering Drawing & Specification Approval Flow
This flow details how new engineering drawings are validated, checked for structural completeness, and indexed.
1. Upload & Categorization
Engineer uploads a drawing PDF and tags metadata: Target Model, Engineering Team, and Category.
2. Quality Checklist Checker (Score Calculator)
The parser scans the text for required blocks: Introduction, Methodology, Results, Conclusion, Safety. The score must be >= 60/100 to pass; otherwise, the file is rejected.
3. QA Approval Queue
Approved checklist drafts go to the QA Approval Queue. A QA Admin must approve, reject, or request changes.
4. Live Index Injection
Upon QA approval, the vector indexes are rebuilt in memory. The drawing becomes searchable by factory operator chatbots instantly.
Vector Retrieval & Stream Generation Flow
Execution Sequence:
- Query Ingest: User posts query to
/chat. - Parallel Matcher: Matches query embeddings against FAQ matrix, approved PDFs, and broad QC guidelines (Mizen Boushi reference database).
- RAG Compilation: Gathers top 5 matches, compile OpenAI prompt containing user message and ground evidence.
- Stream Output: Outputs NDJSON stream back to client (metadata context -> stream token deltas -> done token).
Platform Enforced Business Rules
Threshold: Structural checklist score must be ≥ 60/100.
Explanation: Each drawing must feature critical quality sections (Introduction, Methodology, Results, Conclusion, Safety Warning) to verify model integrity. Lacking more than two blocks halts ingestion to maintain safety data standards.
Threshold: Cosine Similarity Matching score must be ≥ 0.45.
Explanation: Queries scoring below 0.45 are routed to fallback answers. This prevents hallucinated answers or mismatched technical parameters, ensuring users only receive verified answers.
Threshold: Lexical Keyword Matching criteria.
Rules:
- Keywords
drawing, design, engineering, system-> routes to Engineering. - Keywords
production, line, process, manufacturing-> routes to Production. - Keywords
part, inspection, defect, quality, approval-> routes to Parts Quality. - Keywords
new model, drbfm, mizen, gd3, concern point-> routes to QA New Model Development.
User Walkthrough Guide
Customer Support Walkthrough
Step 1: Choose Connected Department
Navigate to the chat screen and select your department from the dropdown. This loads specialized welcome messages and quick action chips.
Step 2: Input Questions
Type a natural language query or click a quick suggestion chip. Wait for the blinking typing indicator to receive a grounded answer.
Step 3: Rate Answer & Escalate
If the response is inaccurate, click Thumbs Down. Fill out the escalation form with your phone and query reasons to alert the QA Admin.
Admin Dashboard Management Walkthrough
Step 1: Authenticate Session
Access the Admin Dashboard. Enter username qa_admin and password QAAdmin123! to acquire authorization credentials.
Step 2: Add FAQ & Re-embed
Go to the FAQ tab. Add questions, answers, and assign departments. Click add to instantly compute new ONNX embeddings.
Step 3: Process PDF Specifications
Go to the Drawings tab. Drag and drop spec drawings. The parser runs checklist scans and updates indices upon passing scores.
Stakeholder Q&A
GAR traces the percentage of user queries that find a high-confidence grounded match in our databases. High GAR shows that support operators retrieve verified, risk-free answers. Low GAR prompts QA teams to write additional FAQs to address new queries.
PyTorch models consume > 500MB of RAM and are slow to initialize. FastEmbed runs MiniLM embeddings on an ONNX pipeline, consuming less than 100MB of memory while returning search matches in less than 35 milliseconds. This makes it highly suited for lightweight, cost-effective servers.
By scanning for sections (like Safety Warnings and Conclusions), the checklist ensures no incomplete engineering documents are ingested. This acts as a quality gate on model data.
Pitch Deck & Presentation Guide
Core Technical Interview Prep
Answer: AutoCare triggers a deterministic lexical hashing fallback. It tokenizes the input text, hashes the tokens using SHA256, and maps them to a normalized 384-dimensional vector. This ensures similarity math continues to function without ML runtime dependencies.
Answer: Upon login, the system creates an HMAC SHA256-signed base64 session token containing role permissions and an 8-hour expiration. Endpoints verify permissions (e.g. approve endpoints require the QA Admin role).