Candidate sessions to fill gaps in the existing AI track. Format options per session: 2-hour workshop or talk/presentation.
Existing track (for reference — don't duplicate): Kickoff Bootcamp (vibe coding), AI & Security, AI Driven Development (Context/Compounding/Harness Engineering), MCP+Skills, MCP Servers, Predicting Mental Fatigue (Pierre), Math Behind the AI Curtain (Tom).
Status: ✅ Selected · 🔭 Scout expedition published 2026-06-04 → Atlas (5 angles: methods, tooling, CI/CD, agent/RAG, workshop design). Format TBD (2h workshop vs talk).
Research verdict: Run as a 2h demo-driven hands-on session standardized on DeepEval (pytest-native, provider-neutral — note OpenAI acquired Promptfoo Mar 2026); teach error analysis, not tool-clicking.
Pitch: The track teaches building agents/MCP/harnesses but never how to prove they work or catch silent prompt regressions. This is the QA discipline of AI — where most teams and clients are flying blind.
Fits because: Direct complement to AI Driven Development. Compounding engineering rots without a regression net; evals are that net.
If 2h workshop:
- Build a golden dataset for a real task (input → expected output)
- LLM-as-judge: write a grader, discover why graders need grading
- Wire evals into CI — fail the build on prompt regression
- Catch a planted regression live
- Possible add-on (from cut #3): read a real agent trace (Langfuse/LangSmith)
- add cost/loop guardrails. Evals tells you if it broke; tracing tells why.
If talk:
- Why vibes don't scale; failure modes of untested AI
- Eval taxonomy (golden sets, LLM-as-judge, rubrics, A/B)
- Tooling landscape (Promptfoo, Braintrust, DeepEval)
- War stories: silent regressions
Alt title: Vibes Don't Scale: Testing the Untestable
Open: format? · tooling (agnostic vs pick one) · presenter?
Status: ✅ Selected — format TBD (2h workshop vs talk)
Pitch: Zero retrieval content in the track, yet "point the AI at our docs/DB and let people ask questions" is the #1 client request. RAG demos trivially and falls apart in prod — wrong chunks, stale data, confident hallucinations.
Fits because: Pairs with MCP (MCP connects tools; RAG feeds knowledge) — the two halves of giving an agent access to stuff it wasn't trained on.
If 2h workshop:
- Embed a real corpus, store in pgvector/Qdrant
- Build naive RAG → watch it retrieve garbage
- Fix it: chunking strategy, hybrid search, reranking
- Add an "I don't know" guardrail to stop hallucination
If talk:
- How embeddings actually work (vector space intuition)
- The RAG pipeline and where each stage breaks
- Naive vs hybrid vs reranked — with retrieval-quality numbers
- When RAG is the wrong answer (long-context / fine-tune instead)
Alt title: Your AI Can't Read Your Docs (Yet)
Open: format? · slot before or after the MCP pair? · presenter?
Status: ✅ Selected — 2h workshop (absorbs the "Local & Open Models" idea)
Pitch: Run the same task through three engines and measure everything. Answers the recurring client blocker "we can't send data to OpenAI/Anthropic" — turns "we can't use AI" into "here's the tradeoff, measured."
Fits because: Unique angle for an EU consultancy (data residency / GDPR). The hybrid lane is the spine — not just "local is worse, news at 11."
Three engines, same task:
- Frontier (Claude/GPT via API) — the quality ceiling
- Local (Llama/Qwen via Ollama) — the privacy floor
- Hybrid — local-first, escalate to frontier on low confidence; or a router (easy tasks local, hard ones up). Often ~90% of frontier quality, most data stays in.
Score all three on one rubric: quality · latency · cost · "could this data legally leave the building?"
Dependency: comes after Evals — reuses the eval harness to score the runs. Clean chain: Evals → Bake-Off.
Keep distinct from: the coding-tool Cage Match (Claude Code vs Cursor vs Copilot) — that compares harnesses, not models. Separate session, see below.
Open: which local models to ship preconfigured? · presenter?
Status: ✅ Selected — talk + live demo (not a hands-on workshop)
Pitch: Same real coding task ("add a feature to this repo with tests") run live through Claude Code vs Cursor vs Copilot vs Codex. Same prompt, same repo, side by side. Who finishes? Who writes tests unprompted? Who hallucinates an API?
Fits because: Sequel to the Kickoff Bootcamp (Claude Code-specific). Bootcamp taught one tool deeply; this zooms out so people choose deliberately, not by habit. Crowd-pleaser in the "Skynet Started?" register.
Why talk not workshop: the comparison only works if everyone watches the same runs. Everyone running their own = noise.
Arc:
- One task, defined upfront with clear success criteria
- Run all four live (pre-recorded fallback — live agents are flaky on stage)
- Score on one rubric: correctness · tests · speed · cost · hand-holding
- Verdict: which tool for which job (greenfield vs legacy, big repo vs script)
Risk to name: dates fast — ranking could flip in 3 months. Frame "as of [date]".
Open: which tools make the cut? · live vs pre-recorded? · presenter?
Status: 📻 On the radar — not scheduling, but worth keeping
Pitch: The Act is in force, phasing in through 2026–27; most teams don't know which systems are now regulated. Pure differentiation — "how EU AI regulation hits the code we ship for clients" isn't on a thousand YouTube channels.
Fits because: Belgian consultancy — geography is an asset here. Bookend to AI & Security: security is "how it breaks," compliance is "what you're liable for." Positions Itenium as advisors, not just builders.
Format: talk (~1h, maybe panel tail). Not a workshop — analysis, not code.
Arc:
- Risk tiers: prohibited → high-risk → limited → minimal, with real client examples
- What "high-risk" obligates: docs, human oversight, logging, conformity assessment
- Transparency rules hitting everyone: chatbots, generated-content disclosure
- Timeline: enforceable now vs 2026 vs 2027
- Practical checklist to run against a client system
Risk: least fun session — no demo, no "whoa." But most commercially useful.
Open: presenter? · external speaker / legal partner?
Status: ✅ Selected — contacted Tom · format TBD (talk + live gen vs workshop)
Pitch: The whole track is LLMs — text in, text out. Image/video gen works completely differently: no next-token prediction, but iterative denoising. Pulls the curtain on how typing words turns random noise into a photograph.
Fits because: Sibling to Tom's Math Behind the AI Curtain — that demystifies LLM internals, this demystifies the other half of modern AI that LLM-focused devs never learn. Wouter runs Midjourney → real practitioner angle, not just theory.
If talk + live gen:
- Forward diffusion (add noise) vs reverse (learn to remove it)
- How text steers denoising (CLIP / conditioning) — what the prompt actually does
- Why hands/text used to break, and what fixed it
- Live: same prompt, watch the denoising steps; then real Midjourney prompt-craft
- Where it's going: video (Sora-style), ControlNet, workflow tools
If workshop: ComfyUI node-graphs / local Stable Diffusion, hands-on.
Honest read: most "fun/curiosity," least "bill a client tomorrow." Enrichment, not capability — but genuinely fascinating and complements Tom's math talk.
Open: talk vs workshop? · pair/co-present with Tom's session?
Status: ✅ Selected — 2h workshop (hands-on by nature)
Pitch: "Let non-technical people just ask the database in plain English." The demo that lights up executives and quietly fails in prod when the model invents a column or table-scans 40M rows. Shows the real, safe version.
Fits because: Most concretely client-deliverable session in the pool — every client has a DB and someone who wishes they could query it without SQL. Applied follow-on to RAG (#2): semantic search over rows is the same embedding machinery.
Arc:
- Naive text-to-SQL: hand it the schema, watch it hallucinate a column
- Make it safe: schema grounding, read-only roles, query validation, row limits, cost guards
- Semantic search: pgvector over a products/docs table — "find similar" not "find exact"
- Hybrid: when to translate-to-SQL vs embed-and-retrieve, and combining both
Overlap risk: shares the embedding/pgvector foundation with RAG (#2). Either reinforcement (this is the structured-data application) or redundancy. Could fold into the RAG workshop's back half if the track needs to be tighter.
Open: standalone vs RAG back-half? · presenter?