Science films, research software and reports in molecular biology and neuroscience, made by leading fleets of AI coding agents. Each piece comes with its proof.
Films & teaching
Each one comes with a way to check it: its sources, a structure or a dataset.
Film · 8 minutes · narrated
How can one forty-minute infusion lift a depression within a day? An investigation from a person down to one trapped molecule, and back.
Short films · a minute or less
One synapse built from real protein structures, showing where an antipsychotic, a benzodiazepine, an SSRI and a psychedelic dock. It opened a series of films that each answer one question in under a minute.
Film · 2 minutes · narrated
How an antidepressant works at the synapse: one serotonin release, run twice, the second time with escitalopram in the transporter.
Scene · 9 seconds · silent
One real nerve cell fires inside a glass mouse brain, and each region it wires into lights up as the signal arrives: 47 centimetres of wiring from a single cell.
Lesson · 17 min · 5 questions
How PET works, and what it has shown in psychiatric disorders.
Lesson · 14 min · 5 decisions
One tracer, followed from a biological question to studies in people.
Lesson · 19 min · 5 questions
How one tracer is used to study noradrenaline, from pigs and rats to patients.
Interactive
What a PET scan can, and cannot, say about synapses.
Interactive · six steps
A receptor in its membrane, rendered from published protein structures (PDB 8TYW, 8T1W).
Trailer · 1 minute · narrated
How the research reports are made: agents read the literature and trace every claim to a real source, and it comes back as chapters, figures and short films. The reports themselves are not public yet.
Film for a research unit · 1:38
Depression is one name for 227 different ways of meeting the same criteria.
Lab film · 0:46
From one detected PET event to the lab's data, methods and themes.
Methods film · 0:48
A brain section builds up, count by count, in real-time autoradiography.
Methods film · 1:00
Why PET in the living brain is paired with autoradiography in tissue.
Website
The website I built for the lab: its research, methods, films and teaching in one place.
In teachingSSRIs: The Replay is used in an MSc course in translational neuropharmacology, and Seeing Chemistry in a course in biological psychiatry, both at Aarhus University.
CommissionedThe one-minute films were made for an international course in psychopharmacology.
Check itThe lab's Watch page lists the courses and each film's sources.
Algorithms & results
Standard bioinformatics tools, rebuilt to run tens to thousands of times faster. Each is a drop-in replacement, tested byte for byte against the original.
| Tool | Replaces | Speed-up | Tested by |
|---|
Seconds per scan · the bars fill in real time
CMRx4DFlow2026, the 4D flow MRI challenge held with MICCAI 2026: reconstruct each scan as fast as possible on one fixed GPU, with no loss of accuracy. Fourteen teams reached the final round. I entered on my own.
How both are made
The films and the software come from one way of working. This is the year it took, drawn from the logs.
Film · 70 seconds · December 2025 to November 2026
Every light is one AI agent, placed on the day it started, and the higher the curve climbs, the more agents had run by then.
molecular biology neuroscience AI
What I bring isn't expertise in MRI or PET. It's leading a fleet of coding agents.
I design the system around them: the harness they work in and the loops that verify every step. I set up teams of agents that check each other's weaknesses. And when they hit diminishing returns, I get them unstuck.
about 1:10 · silent
A one-minute explainer of the MRI entry, one real run followed through, goes here.
Before any agent starts, the task gets a test it can't argue with. For SwiftX: the output must match the original tool, byte for byte, on every test.
The work is split into pieces, and agents take them in parallel inside that harness. of them so far.
Other agents review each piece and look for the weaknesses its author missed. Nothing is accepted on its author's word.
When progress flattens, I step in with a new angle, a different check or another team.
The long list
pieces from May to November 2026, newest first: finished films, single scenes, pages, papers and studies along the way. Point at a tile to play it.
A piece is dated by its release when it has one, and otherwise by the day its file was made. The agents' work starts in December 2025; the catalogue of finished pieces starts in May.
I'm a master's student in molecular biology at Aarhus University, doing my project at the Translational Neuropsychiatry Unit.
I started making these films and tools in my own time. What they share is how they are made: with fleets of AI coding agents, and with a check on every step, so that anyone can see what a piece rests on.
If you'd like to work together, or want to know how something here was made, write to me.
joe@joechristensen.net