KARL

KARL is the cloud-intelligence system I built end to end inside Orange Business's Cloud Avenue product team. It started as an automated competitive-intelligence robot, in production since July 2025: automated collection and syntheses sent to the teams, first orchestrated with N8N. That robot is what I demoed live to Aliette Mousnier-Lompré, CEO of Orange Business, in July 2025 (documented in a public LinkedIn post), and its usage was discussed at executive committee level in August 2025. In early 2026 I paused it to bring it into compliance, then rebuilt it entirely: N8N replaced by LangChain + ChromaDB, with models served locally (Llama 3.3 70B, DeepSeek R1, QwQ 32B) via vLLM on H100 NVL and L40S GPUs, for compliance and sovereignty. KARL came back to production in that form, with a conversational RAG agent on top of the intelligence feed. The goal was never a demo that impresses: answers grounded in internal sources, auditable and reliable, for product and sales teams.

Automated intelligence feed

fig. 02 · Illustration. From automatically collected sources to the synthesis sent to the teams

A cloud competitive-intelligence robot replaces manual monitoring: automated collection and syntheses sent to product and sales teams, in production since July 2025. The first version was orchestrated with N8N.

Rebuilt on local GPUs

fig. 03 · Illustration. Models served via vLLM on H100 NVL and L40S, ChromaDB vector store, conversational agent

In early 2026, a deliberate pause for compliance, then a full rebuild: N8N replaced by LangChain + ChromaDB, with Llama 3.3 70B, DeepSeek R1 and QwQ 32B served locally via vLLM on H100 NVL and L40S GPUs. Data stays in-house.

Grounded, auditable answers

fig. 04 · Illustration. Every RAG answer points back to the internal documents it rests on

On top of the feed, a conversational RAG agent grounds its answers in internal sources. The goal was never a demo that impresses: auditable, reliable answers, evaluated against real cases instead of gut feeling.

Challenges

• Going from manual competitive intelligence to reliable automated collection and syntheses
• Bringing the system into compliance without killing it: an assumed pause in early 2026, then a full rebuild
• Serving multiple LLMs locally on GPUs (H100 NVL, L40S) via vLLM to keep data in-house
• Getting auditable, reliable answers rather than ones that just look good in a demo

Solutions

• First version orchestrated with N8N: automated collection and syntheses sent to the teams
• Full rebuild on LangChain + ChromaDB, local models (Llama 3.3 70B, DeepSeek R1, QwQ 32B) served via vLLM
• RAG pipeline to ground answers in internal sources
• Evaluating outputs against real cases instead of trusting vibes

Results

• Demoed live to Orange Business CEO Aliette Mousnier-Lompré in July 2025 (public LinkedIn post)
• Usage discussed at executive committee level in August 2025
• Intelligence feed in production since July 2025, reborn after a full rebuild (N8N to LangChain + ChromaDB + local GPUs)
• In production for Orange Business product and sales teams

LinkedIn post

Technologies

LangChain · ChromaDB · vLLM · H100 NVL · Llama 3.3 70B · DeepSeek R1 · RAG · N8N · Python