AI-Powered Workflow Automation

Context

TECHWAVE is an industrial electronics group. During a 7-month internship (Dec 2024 – Jul 2025), I worked as the interface between the AI consultancy OPEO and TECHWAVE's engineering and procurement teams — translating operational pain points into deployable Python automation pipelines.

All pipelines run entirely on-premise — no data leaves the company network. This was a hard requirement given the confidentiality of client specifications and regulatory documents.

Impact

Across 5 use cases, the pipelines deliver:

| Use case | Time without AI (days/year) | Time with AI | Time saved | |---|---|---|---| | Order processing | 13.9 | 1.96 | 11.94 | | RAO sourcing | 12 | 1.71 | 10.29 | | BOM update (Refab) | 7 | 1 | 6 | | Traceability & review | 40 | 5.71 | 34.29 | | Compliance matrices | 46.25 | 6.61 | 39.64 | | TOTAL | 119 days | 17 days | 102 days/year — €85,750/year |

Pipeline 1 — BOM Sourcing via Mouser/DigiKey API

Given a Bill of Materials (BOM) from a client file, this pipeline:

1. Detects the header row automatically and maps column names intelligently
2. Queries the Mouser Electronics REST API for pricing, availability, and lead times per reference
3. Exports a formatted Excel report ready for import into Sage 100 ERP (BLC module)

Result: eliminated ~12 days/year of manual component searches.

Pipeline 2 — Regulatory Requirement Extraction

This is the most technically complex pipeline. Input: regulatory PDF documents (CDC, CGA — often 200+ pages). Output: structured compliance matrix with article, description, and technical/legal classification.

Three approaches compared:

1. Method 1 (no AI): requirement = section title + paragraph — fast but imprecise
2. Method 2 (LLM): full LLM extraction — flexible but prone to hallucinations on dense regulatory text
3. Method 3 (Hybrid): PyMuPDF parses structure deterministically; LLM only handles ambiguous blocks → chosen for production

The hybrid approach was validated on real CDC and CGA documents from TECHWAVE clients.

Local LLM Infrastructure

All LLM inference runs locally via Ollama (llama3 model), wrapped in a TECHWAVE-branded secure chat interface. The system:

• Accepts file drag-and-drop for document processing
• Encrypts connections end-to-end (displayed as "Chat sécurisé & confidentiel")
• Stores no data externally

What This Demonstrates

• Production-grade Python: modular pipelines with error handling, column matching, and formatted outputs — not notebook code
• Local LLM deployment: real-world confidentiality constraint → architectural decision (on-premise inference)
• Hybrid pipelines: combining rule-based reliability with LLM flexibility, rather than blindly using LLMs for everything
• Cross-functional impact: understanding procurement, compliance, and engineering workflows well enough to automate them

---

Client names, specific document content, and detailed metrics are confidential. Figures shown with permission.