Video summary

Il a créé une IA pour faire son travail (son boss hallucine)

Main summary

Key takeaways

Technology

Tech summary (AI automation system, tools, and agent workflow)

  • Core premise: Sami demonstrates how he automates parts of his real work using a team of specialized AI agents running 24/7—claiming it handles ~50% of his tasks. He argues automation is possible even if you’re not initially an ML/AI expert, because the key is building the right workflow and tools.

Local / privacy-first approach

  • He avoids mainstream hosted LLMs (e.g., sending company data to GPT) due to data security uncertainty.
  • He runs models locally on a custom ~€5,000 PC, including experimentation with a crypto-mining GPU rig / multi-GPU motherboard (“RIC”-style).
  • He uses self-hosted tooling such as Ollama-like model management and AnythingLLM (Docker-hosted) to create isolated workspaces, so departments don’t share memory/data across conversations.

Model/data orchestration

  • RAG vs. fine-tuning: He discusses RAG (Retrieval-Augmented Generation) as an alternative to heavier fine-tuning, claiming RAG often yields equal or better practical results while being operationally simpler.
  • Single model vs. multi-agent: He contrasts:
    • sending context to one big model/chatbot, vs.
    • using a multi-agent setup that effectively splits context to fit around context limits.

Example productized automations and “levels” of projects

Level 1 — Internal intranet automation (dashboards + order handling)

  • His first meaningful automation milestone improved an intranet system used by maintenance staff and for administrative workflows.
  • Goal: when an agent places an order, managers use dashboards to verify and decide using data aggregated “from everywhere.”
  • He notes common software-dev bottlenecks (UX/UI + coding + costly agency/ESN work), including cases where agency development exceeded €100k.
  • He automates the backend/software generation using an agent team structure.

Level 1.5 — Google Ads keyword pipeline via agent teams (CrewAI)

He built an agent workflow to automate keyword research and placement for a Google Ads network:

  • A research assistant agent crawls Google and competitor top positions, producing an Excel list.
  • A content strategist agent checks whether competitor landing pages match the intended customer type (B2C vs B2B), filtering keywords accordingly.
  • A manager/supervisor (prompt-driven) oversees quality checks; if CSV formatting is wrong, it re-runs.
  • Agents transform CSV → paid request → Google integration (he references using LangChain and Google connectivity).

He says it originally ran fully automatically, but later added manual activation for Google campaigns after testing.

Level 2 — Microsoft/Active Directory sync via local AI-assisted coding

  • The environment is Microsoft-based, centered on Active Directory (users/phone numbers/titles).
  • Instead of buying an external sync service (citing ~€15/month/user and estimating ~€5,000/year), he built his own:
    • a phonebook/directory synchronization application (using Kotlin for Android),
    • with help from Cursor (an AI coding assistant, a VS Code fork) that can edit multiple files across a codebase within a single discussion.
  • He claims a working service in about 2 hours, contrasting this with third-party costs.

Level 3 — Rebuilding an entire React intranet with “mega prompt” + agent prompt manager

After a previous agency rewrite cost ~€100k, he recreated functionality from scratch using:

  • A “vision model” to describe screens (he references Pixel/Sigma-related tools in the subtitles).
  • Turning each screen into prompts, then compiling them into a large Markdown “mega prompt” describing components.
  • Agents implementing the system; it took about 3 days to generate the React project.

He explains why a single chatbot struggles (context window limits) and why multi-agent approaches help:

  • He uses a “prompt manager” that segments and routes parts of the prompt to smaller agents.
  • Each agent receives only relevant compressed global context, reducing errors and improving specialization (front-end vs back-end vs DevOps-style roles).

Level 4 — iOS + Bluetooth reverse-engineering app for an electric bike (personal app)

Sami describes building a native-feeling app for an electric bike brand (Super73):

  • Official apps required internet, accounts, and could be restricted remotely (features could be removed).
  • Alternate apps exist (often via reverse engineering the controller), but he wanted custom features plus native mobile integration.

Two-team structure

  1. Reverse engineering team
    • Discovers controller-accessible Bluetooth routes (button actions/modes).
  2. App-building team
    • Implements features and UI using the route documentation.

Tools and workflow

  • He mentions a Python tool that tests routes and generates a CSV of routes that respond successfully (manual Bluetooth testing is still needed).

Demonstrated features

  • Auto-connect UI, nearby bike detection/proximity (“squads”), and mode switching (including a previously restricted mode)
  • Trip recorder, GPS navigation with bike-optimized routes
  • “Danger rating” for roads using external data/APIs about cycle paths
  • Avoid certain routes based on weather and battery
  • Gamification/achievements

Reported results and next steps

  • Early iOS version: about a weekend
  • Revenue: crossed ~$10,000 in 2 months (pricing mentioned: about $10 per app)
  • Next steps:
    • Added Apple Shortcuts / Siri integration via Apple shortcuts
    • Planned Android expansion by converting functionality
    • He credits agent-based code translation (no Cursor used during Android conversion) and says it resulted in a native Android Kotlin Bluetooth app
    • Estimates Android conversion took about a week, with some back-and-forth

Reviews / warnings / limitations (as stated by Sami)

  • Primary constraint: Even if automation feels easy afterward, there’s a learning curve during initial setup.
  • Productivity impact: He believes current tech is sufficient for businesses now; it can boost productivity, potentially reduce team size but won’t eliminate developers.
  • Testing challenges: He tried dedicated “test reviewer/bug hunter” agents but couldn’t fully get them working as expected (at least at the time described).

Sponsors / side tool mention (non-technical but included)

  • Swiss Transfer: Mentioned for sending large files (up to 50GB), with password protection, expiration, download limits, and storage in Switzerland.
  • Unea: A collaborative email translation/rewrite feature integrated into their suite, described as “sovereign” with data processed under European law.

Main speakers/sources

  • Main speaker/guest: Sami
  • Mentioned tools/projects/sources:
    • AnythingLLM, Ollama (implied), LangChain, CrewAI (and the agent-team approach)
    • RAG, Cursor (VS Code fork)
    • Active Directory / Microsoft
    • Python (route testing)
    • Google Ads
    • Bluetooth APIs
    • Apple Shortcuts / iOS
    • Kotlin / Android
  • Sponsor/source: Swiss Transfer (via Infaniaac in subtitles)

Original video