Automation, Optimization & AI
Freelance in Lyon: no-code automation (n8n, Zapier) or custom-built (Python, APIs), office automation, reporting, and AI agents for triage and assisted drafting. Two real case studies in detail.
What I do
No-code or custom-built automation
n8n, Zapier when speed matters and the need is standard. Python and APIs when no-code hits its limits. The right tool for the need, not the trendy tech.
Office automation
Outlook VBA for sorting, filing, and automatic follow-ups. Advanced Excel to make processes that were still run by hand more reliable.
Reporting & data visualization
Datastudio to turn scattered data into readable, actionable dashboards.
AI agents
Building LLM agents for triage and assisted drafting — not just standard no-code, but agents that make decisions on real cases.
Why my track record proves it
Technical flexibility doesn’t show on a résumé, it shows on real cases that are very different from one another.
I built a data pipeline that aggregates, deduplicates, and summarizes heterogeneous sources. I also built a triage and assisted-drafting agent on a sensitive subject — banking emails, where a misclassification has a real cost.
Two problems of a completely different nature, two different solutions. That’s what the ability to go from no-code all the way to custom development, depending on what the situation actually calls for, looks like — not applying the same recipe everywhere.
Track record
Newsletter aggregation and summarization pipeline
Gmail · LLM agent · Data table · Web
- The real problem wasn’t the number of newsletters received, but the repetition: the more sources you add, the more you re-read the same information from different angles or with different numbers.
- A pipeline connecting 2 source Gmail accounts to an LLM agent (extraction + summarization), storing results in a data table under a unique constraint on the source URL, published on a dedicated web front-end (newsletter.edgar.center).
- Real obstacle: deduplication by URL alone wasn’t enough — the same news sometimes circulates under several different URLs or phrasings. The summaries were also poorly calibrated at first (an in-depth article treated like a short brief, which lost exactly the information the tool was meant to save time reading). Both problems were fixed.
2 source Gmail accounts · deduplication and calibration fixed · 6 topics tracked without repetition (Finance, Politics, Tech, AI, robotics, French industry)
Triage and assisted-drafting tool for banking emails
LLM agent · Triage · Assisted drafting
- Each client being supported had about ten connected banks, each generating daily email exchanges. Volume grew steadily with the number of clients until it became unmanageable by hand: 3 people for around 150 emails/day, a pace already at its limit with no room for growth.
- An LLM agent classifies each email (concerns the team vs. a simple client↔bank exchange) and drafts a reply for genuine questions.
- Real obstacle: the team’s initial distrust of AI-generated drafts — a sensitive subject, with banks at stake. No shortcut: systematic human validation of every draft at first, trust was built over time as the rate of good drafts was observed.
300 emails/day processed · 3 people (150 emails/day) → 1 person (300 emails/day) · output per person multiplied by 6