The Problem
After my 23rd attempt at debugging an inference pipeline at 2 AM on a Sunday, I realized most ML tutorials end exactly where real problems begin.
My Path Here
I didn’t come from a traditional CS background — I studied economics and law. But curiosity (and a lot of debugging in PyTorch) led me into machine learning.
Over the last 5+ years, I’ve worked on:
- Evaluation frameworks such as Dynabench.
- Large-scale datasets such as People’s Speech and Adversarial Nibbler.
- Coauthoring and publishing at top conferences:
- HelpMed, at Nature Medicine 2026.
- Constrained Wikigame, at the Reasoniing LLMs Workshop at ICLR 2026.
- Unsupervised Speech in The Wild Challenge,
- Best paper award at NeurIPS D&B 2024 with The Prism Alignment Dataset.
- BabyLM at CoNLL (ACL) 2023.
- DataPerf at NeurIPS 2023.
- Competitions every now and then, such as PatentBot.
Through this work, I noticed a consistent pattern: the gap between research papers and production reality is bigger than anyone admits.
- Technical content: often too dense for anyone outside the research bubble.
- Popular ML content: often oversimplified to the point of being misleading.
- Bridges between the two: almost nonexistent.
That’s the problem this blog tries to solve.
Two Tracks, One Mission
This blog runs on two parallel tracks — each aimed at closing the gap between research depth and real-world clarity.
For Engineers
Deep technical dives, reproducible experiments, and hard-won lessons from production. Sample posts:
- Why our LangChain pipeline breaks when trying to customize it
- Scaling inference when GPUs are scarce (and expensive)
- Fine-tuning small LLMs for common use cases — a cheaper path than you think
- Building a production-ready ASR pipeline: from VAD to deployment
For Everyone
Clear, accessible explainers of how ML shapes daily life — no PhD required, no jargon walls. Sample posts:
- Why LLMs hallucinate (and why it matters in practice)
- Getting more out of your AI assistant without burning tokens
- Tips for better image generations with today’s generative models
- How recommendation systems really guess your taste
Why “The Artificial Engineer”?
I’m not a “real” engineer by degree, but I’ve spent some time making AI systems work in the real world. The name captures the irony: someone artificial (by training) engineering artificial intelligence.
Join the Conversation
I’ll be posting often, hopefully. But I want this to be a dialogue, not a broadcast.
