Hi there,

Welcome back to Beyond the Prompt! I hope you're having a wonderful start to the Spring season. Over here, we kept pretty busy... in Belgium, France, and Luxembourg... but also in Canada, where we're working on interesting opportunities... More on that in a couple of months if all goes well...

🤗 Before we start: we're hiring!!!

Besides the usual Data Science and Engineering gigs, we have two exciting opportunities open right now

• An Administrative and Finance Manager (part time) to help me steer the ship in an efficient way.
• A Data Engineering Manager to grow this thriving practice alongside our brilliant experts.

💡 From the case studies

Roadmapping in Banking. Recent advancements in AI are prompting many people to reassess their priorities because of significant changes in what is technically possible. Needless to say, we're more than capable to help.

📜 From the blog

• No need to be fast nor furious to talk about drift, the data science kind.
• We have a nice write-up of Christophe's podcast appearance, if you prefer the textual version.

🎙️ From the podcast

After Chris last month, this time it's Yours Truly's turn to be in the hot seat. If you're less tired of hearing the sound of my voice than I am, head this way for my nuggets of wisdom (NB: the podcast is in French).

🧪 From the lab: What we're cooking

Graph RAG, enhancing Information Retrieval: we worked on comparing standard RAG (Retrieval-Augmented Generation) and Graph RAG approaches. Standard RAG uses vector similarity searches for document embeddings, but Graph RAG adds graph relationships to enhance retrieval accuracy in large and scattered document collections. Key insights include:

We benchmarked TimeGPT, a foundational model for time series forecasting, against traditional models like XGBoost.

We tested Power Automate's integration with Dynamics 365 and found that automated cloud flows can efficiently streamline processes, making it a good choice for small projects.

Copilot Teams Bot for faster knowledge retrieval. We're testing a bot integrated with MS Teams to access past project information from our internal Projects database in Notion.

🌐 From everywhere else

ChatGPT is getting a little less sycophantic. Good, because its kiss-ass tone was getting on my nerves...

Stop being polite with your GPTs, it's costing millions.

Speaking of millions, Ed Zitron has doubts about the financial viability of OpenAI...

In a memo that might become part of tech history, Shopify's CEO mandates employees use AI before requesting extra resources. I agree. AI tools must be part of one's toolbox to get unstuck and structure thoughts. Sure, they have plenty of limitations and can't be used on their own, but they provide too much added informational value to be ignored today.

As AI practitioners, we have a duty to be part of the solution and not part of the problem. A couple of cautionary tales popped up in my feed recently.

• Brazil's AI social security app, while aimed at reducing bureaucracy, struggles with complex claims, leading to automatic denials that disproportionately affect vulnerable rural workers.
• HP introduced, then subsequently ditched a mind-boggling 15-minute wait time policy in some of its call centers, hoping it would steer customers to digital (and AI-powered) channels.

If you haven't watched a Geoguessr expert at work, watch one. Then marvel at the fact that OpenAI's o3 model has become incredibly good at it too.

"Vibe coding" is getting a bad rap at the moment. It's justified when people think it can reply entreprise-grade development. But there are use cases where prototyping where it's a great idea. Deeplearning.ai has an interesting training on Replit.

Can there be a Beyond the Prompt without an article on document management? Not yet. 😇 → Gemini 2.5: The First LLM That Actually Understands PDF Layouts.

What Google giveth in terms of credibility, Google taketh it away... the ad behemoth is scrapping its plans to phase out cookies.

A bad mic in your video calls might be hurting your career prospects.

Finally, does Claude have values? Anthropic's latest research reveals insights into how their AI model, Claude, expresses values in real-world interactions, highlighting both successful alignment with human preferences and areas needing attention.

🧑‍💻 Life @ Agilytic

April is an important month for us. It's the time when we plan a team building trip that coincides with the first rays of sunshine. This year we went to Disneyland Paris, and the team had a blast.

We held a mid-year business update to transparently share the organization's status, our direction, and the larger context. We are pleased with our progress halfway through the fiscal year, considering the current business uncertainty and volatility. The way I sum up our mood is: We earned the right to be optimistic, but we can't take anything for granted.

Lastly, April is also the month when we perform our yearly reviews with the entire team. Regardless of the "grade" the evaluation yields, a good evaluation is one where there are no surprises on either side. For that, we work year-round on continuous feedback and mentoring. It’s definitely not perfect, but I really appreciate how our small team is able to keep up this level of care.

👋 Parting thoughts

As the spring flowers burst into bloom, may your ideas—and your models—flourish even faster. After all, whether you’re fine-tuning an LLM or planning a team-building retreat, continuous feedback is the secret fertilizer. Keep your prompts open, your curiosity in full sunshine, and let’s watch what grows—beyond the prompt!

All the best,

Julien

_

PS: Feedback is a gift. Tell me what you liked (and liked less) simply by replying to this email.