Blog
Field notes.
Healthtech AI, engineering, and applied data science. What works, what quietly breaks, zero hype.
2026-05-24
Carbon: The DNA Foundation Model That Makes Evo 2 Accessible to Biotech SMEs
Two months ago Evo 2 proved AI could read the genome. This week Carbon proved you no longer need a GPU cluster to use it. Carbon-3B matches Evo 2-7B on seven benchmarks with half the parameters and 150× faster inference.
Read2026-05-10
A 95% Accurate AI Model Can Still Be Fragile in Regulatory Review, Here's Why
A pharma AI model can post 95% accuracy and still be fragile under regulatory review. The cause is rarely the metric, it's the absence of a pre-specified document. With ICH M15 reaching Step 4 on January 29, 2026, the Model Analysis Plan is becoming non-negotiable.
Read2026-05-09
ROI-First vs Tech-First: Why Your AI Projects Never Reach Production
Most pharma AI projects never reach production, and the cause is rarely technical. The classic scenario starts with a technology request (RAG, agent, fine-tuning) without quantifying what the system should return. Three variables are enough to flip the order.
Read2026-04-29
When Your Labels Lie: What a Neuroscience PhD Taught Me About Healthcare AI
I plateaued at 73% for months on a Sharp Wave Ripple classifier. The model wasn't the problem, the temporal labels were, contaminated by biology itself. A self-supervised pipeline pushed performance to 84% on the same architecture. In healthcare AI, this pattern is everywhere.
Read2026-04-28
95% Accuracy That Drops to 58%: Why Deep Learning Models Fail on Biological Signals
A 95% accuracy model collapses to 58% when the experimental setup changes, I've seen this play out on neural signals. The issue runs across deep learning on biological data: ECG, EEG, sensors, imaging. The model isn't learning biology, it's learning the signature of an acquisition session.
Read2026-04-22
"Tell Me About Yesterday Morning", The Question That Changes an AI Audit
In an AI audit, job descriptions paint an idealized role, clean, logical, fiction. One question surfaces the real tasks, invisible workarounds, and quantified savings: "Tell me about yesterday morning."
Read2026-04-19
72 Hours, 3 Announcements That Reshape AI Pharma
In 72 hours, Novo Nordisk, Amazon, and OpenAI all entered AI drug discovery. Three takeaways beyond the headlines: Alphabet's monopoly is challenged, Big Cloud targets early-stage biotech, and Big Pharma that delays won't miss one train, they'll miss three.
Read2026-04-19
Funding Your AI Project in Southern France: Up to 4 Stackable Programs in 2026
In 2026, an SME in Southern France can stack up to 4 public funding programs for its AI project, including one fully covered. Only 24 companies have used it so far. Here's a breakdown of available programs and how to cut your bill by three.
Read2026-04-14
AI Is Transforming Spike Sorting, But It Won't Replace the Electrophysiologist
Kilosort4, SimSort, SpikeInterface: AI is reshaping electrophysiology. What took days of manual spike sorting now takes minutes. But algorithms can't judge slice viability or series resistance drift, that's where dual expertise matters.
Read2026-04-11
LangChain or Direct API? The Real Decision Criteria
LangChain is a great framework, but it's a tool, not a strategy. The real decision criterion isn't which framework to use, it's your workflow complexity. Here's when to use LangChain/LangGraph and when a 50-line Python script is the better choice.
Read2026-04-11
France Is Building Its Own AI Infrastructure, What It Means for Your Business
France is deploying SIAAG, a cross-ministry generative AI platform hosted on SecNumCloud. With Albert API, Mistral models, and 10,000 public agents already testing it, this sovereign infrastructure is reshaping requirements for companies working with the public sector.
Read2026-04-08
AI × Pharma Consulting in France: A Confirmed Strategic Gap
I mapped every AI consulting player in pharma across France, strategy firms, IT consultancies, deeptech startups, freelancers. Against 30-50 specialized players in the US/UK, France has zero. Full analysis of a strategic gap.
Read2026-04-04
AI Works Precisely When Data Is Scarce, 3 Proof Points from Rare Diseases
TxGNN, Healx, and Unlearn.AI prove that AI delivers results precisely where data is scarce. Zero-shot drug repurposing, digital twins, and adaptive designs were built for rarity, not despite it.
Read2026-04-01
I Was Coding Backprop by Hand in 2018, Here's What It Taught Me About AI in 2026
In 2018, GPT-1 went unnoticed and deep learning was a niche skill. After completing Andrew Ng's specialization, I learned that understanding fundamentals and stacking skills matters more than chasing trends, a lesson that still holds in 2026.
Read2026-03-30
You Use AI? Here's What the EU AI Act Means for Your Business
If you use ChatGPT, AI-powered CRM scoring, or automated resume screening, you're a "deployer" under the EU AI Act. Starting August 2026, you have concrete compliance obligations, here's what to check now.
Read2026-03-26
Nearly 2 Out of 3 AI Projects Fail to Deliver
IBM, McKinsey, and Gartner agree: most AI projects fail. Not because of technology, but because of 3 recurring mistakes, starting with tech instead of the problem, ignoring data foundations, and underestimating the human factor.
Read2026-03-26
Google's AI Co-Scientist Solved in 48h What Took Researchers 10 Years
Google's AI Co-Scientist, built on Gemini 2.0 with 7 specialized agents, reproduced 10 years of research in 48 hours. Lab-validated results published in Cell and Advanced Science. But 0 FDA approvals so far, here's the reality check.
Read2026-03-22
A Pipeline of 3 AI Agents Can Generate a Scientific Report from A to Z
A 3-agent AI pipeline, reader, writer, quality controller, can automate scientific report generation in minutes. The key is workflow design, not technology.
Read2026-03-18
Evo 2: the AI that reads, understands and writes DNA
Evo 2, published in Nature in March 2026, is the largest biological foundation model ever built. Trained on 9 trillion DNA base pairs, it predicts mutation impact, generates functional DNA, and paves the way for programmable biological design.
Read2026-03-14
3 Mistakes SMEs Make When Talking About AI
ChatGPT is not a strategy, technology before problem-solving is a trap, and your data quality matters more than your model. Here are the 3 most common AI mistakes I see in SMEs.
Read2026-03-11
From the Human Brain to AI: My Journey
12 years in neuroscience research, then AI. Why I launched SG AI Solutions to help SMEs embrace AI transformation.
Read