Digital AI Solutions — Europe
Our monitoring said all systems healthy while an AI agent silently stalled overnight. Here is how we rebuilt health checks for LLM workloads that fail in ways HTTP 200 cannot see.
Hardening a server-rendered platform with Content-Security-Policy looked like a one-line header change. It took down payment flows in three browsers and taught us a rollout pattern we now reuse.
We shipped an LLM-as-judge eval pipeline for a grounded assistant and trusted it for weeks. Then we hand‑labelled 200 cases and found the judge was wrong four times out of ten. Here is how we fixed it.
Pure vector search scored 61% on queries containing SKUs and error codes. Adding BM25 alongside pgvector lifted retrieval accuracy to 89% without touching the embedding model.
A 14-job pipeline looked fine in staging. In production, fan-in ordering cost us a 6am client deadline and taught us three rules we now enforce.
After three months logging every FAQ, HowTo and Product schema on client sites, the citation lift came from three types nobody talks about.
Three agents retried the same failed job in lockstep and cost us €40 in 12 minutes. Here is the jittered, state-aware retry pattern we shipped instead.
A p95 latency spike on a customer-facing agent forced a brutal timeout decision. Here is the data, the trade-off, and what we measured before and after.
Across four SME platforms we rebuilt as server-rendered apps, time-to-interactive dropped by roughly 60%, hosting bills halved, and one whole class of bugs disappeared. Here are the numbers and the trade-offs.
We turned on prompt caching for a grounded assistant and the bill tripled before we noticed. Here is the budget rule we now enforce on every LLM copilot.
Our RAG system quietly degraded as the corpus crossed half a million embeddings. Here is the exact threshold, the metric that caught it, and the fix that did not cost us a rewrite.
Retries are easy to add and brutal in production. How one deduplication key per job stopped duplicate invoices, double-sent emails and a very awkward client call.
We instrumented llms.txt on nine client sites for 90 days. The honest numbers: crawl rate, citation lift, and the bots that never showed up.
An autonomous content agent got stuck in a retry loop and burned through its monthly budget in 48 hours. Here is the guardrail stack we built to make sure it never happens again.
Tatano sells in Luxembourg, Belgium, France and Switzerland. In six months: four country domains, six languages, a 24/7 AI chatbot, a daily autoblog and generative video campaigns — all operated autonomously.
TikTok, Instagram, Facebook, YouTube — four posts a day, every day, across all platforms. No editor, no scheduler, no agency. Here is the architecture that makes it run.
Every night at 03:00 UTC, discover_daily_topic.py finds the hottest controversial AI story (<48h) via Google Trends, YouTube audience signals and Grok web search — then produce_daily.py turns it into a full episode on @lepodcastia without human intervention.
An article a day, published automatically, indexed within minutes, distributed across every channel. The architecture behind a content engine that runs without a content team.
795 MB of licensed trance tracks. Two YouTube channels. TikTok, Facebook, Instagram. A hook bandit algorithm. All driven by TikTok Studio and the Blotato API — zero manual publishing.
Optimus answers product questions 24/7 across tatano.lu, tatano.be, tatano.fr and tatano.ch in French, English, Dutch, German, Italian and Luxembourgish. 1.3-second response time. Here is how it was built.
A premium showcase for an independent Belgian strategy consulting firm — editorial design, sub-200ms TTFB, zero technical debt at launch, and an AI assistant powered by NVIDIA NIM running on the same server.
One Next.js platform. One article per day, auto-generated. Newsletter, social and IndexNow on every publication. GEO audit from 66 to 90+. Zero hours per week of content operations.