Boring infrastructure for un-boring products
Strange Loop, Berlin. On why every interesting AI product is built on deeply boring infrastructure, and why the frameworks that try to hide it are lying to you.
Talks · Keynotes · Workshops · 2018–present
Saying it out loud.
Conference talks, keynotes, and the occasional podcast episode — mostly about conversational AI, shipping ML in production, and why enterprise software is more interesting than it looks.
Latest talk
Boring infrastructure for un-boring products.
Strange Loop · Berlin Apr 2026 42 min
A keynote arguing that the most interesting AI products in 2026 sit on top of the least interesting infrastructure imaginable — and that this is not an accident. Covers queues, retries, schema contracts, and why every "agent framework" eventually becomes a worse version of Kubernetes.
Strange Loop '26№ 12
Boring infrastructure for un-boring products.
A. Nichol · Keynote42 min
01 2026 2 talks
02 2025 4 talks
The 10,000-hour chatbot
AI Engineer Summit, San Francisco. What a decade of watching people ship chatbots has taught me about where they break.
Open source and the enterprise long tail
OSS Europe, Bilbao. The quiet truth that most open-source adoption in the enterprise is driven by procurement, not by engineers.
Shipping against a moving model
AI.Engineer workshop. A hands-on session on building eval pipelines that survive quarterly model updates.
Where did all the intents go?
NLP Summit panel on the shift from classifier-based dialog systems to LLM-native ones — with some pushback.
03 2024 3 talks
Four failure modes of LLM products
QCon San Francisco. The four most common ways LLM products quietly stop working in production, and what to do before they do.
Finetuning is for when you've run out of better ideas
AI Engineer World's Fair. A polemic on finetuning's place in the modern stack — specifically, further down than you think.
The dialog manager comes back around
DevNexus Atlanta. Why the architectural idea we called "dialog management" in 2017 is quietly being rebuilt as "agent orchestration" in 2024.
04 2018–2023 3 talks
Conversational AI, 2013–2023: a trajectory
Web Summit. Ten years of conversational AI, told as a story about three wrong turns and two correct ones.
Against single-turn benchmarks
EMNLP industry track. Why the benchmarks we had were actively misleading us about what users experienced.
Supervised dialogue learning, from scratch
PyData Berlin. The talk that eventually became Rasa Open Source 2.0 — and my first public attempt at explaining what we were building.
Speak at your thing?
I'm selectively available in 2026 for keynotes, workshops, and occasional podcast episodes. Probably a yes if you're a conference about practical ML, an engineering org running an internal event, or a podcast with a guest list I admire.