About · who I am
I’m Andrew — I work at the seam between chemistry and computation: teaching machines to understand molecules, proteins, and the physics underneath them.
Right now I’m an AI intern at Soley Therapeutics, working on molecular machine learning for drug discovery — explainability, generative chemistry, and foundation models for molecular property prediction. Before that, three years in Brian Stoltz’s lab at Caltech: total synthesis at the bench, computational modeling in ORCA, and software to automate the lab around me.
I’m a polymath by temperament. I’ve shipped apps used by tens of thousands of people, reverse-engineered a lab robot’s firmware until it played Liszt, and turned prime numbers into music. The throughline is simple: I want to solve the hardest problems in science, and I think computation is how we get there.
Research & lab
Before the machine-learning work, I did total synthesis in Brian Stoltz’s lab at Caltech — making complex natural-product molecules from scratch, characterizing every intermediate, and writing software to automate the instruments around me.
Total synthesis
Three years in a Caltech total-synthesis lab, mostly at the bench. I completed 16 consecutive reaction steps toward a complex diterpenoid intermediate, plus shorter fragment syntheses on the side — each step purified, characterized, and carried forward. Roughly 1,000 hours of hands-on organic synthesis.
16 consecutive stepsSeparation & analysis
The daily craft of synthesis is separation and identification: column chromatography and HPLC to pull one compound from a mixture, then LC/MS, GC/MS, and NMR to prove what you made. When an HPLC autosampler part broke, I modeled and 3D-printed a replacement — about $250 saved per part.
3D-printed HPLC partComputation & automation
I reverse-engineered a liquid-handling robot’s M68k firmware in Ghidra to drive it from code — and, to prove I’d cracked the protocol, made its motors play Liszt’s “La Campanella.” A Python-over-serial bridge piped live reaction status into a Zoom chat so I could watch overnight runs from anywhere; another tool pulled 2,600 catalysts from a supplier’s API into a searchable catalog. For the chemistry itself: ORCA DFT to rationalize a key step’s selectivity, and AutoDock Vina docking against an AlphaFold-modeled receptor.
Ghidra · ORCA · VinaBeyond the lab
At the piano
I’ve performed at Carnegie Hall and Merkin Hall. This is my Étude Op. 10 No. 12 in C minor — rendered live from the score, every note in real time. Tap the speaker in the corner to hear it.
Watch me play ↗MIDI: B. Krüger · piano-midi.de (CC BY-SA) · composition public domain
Curiosities
Speedcubing — 9.42s, and solved blindfolded. 2×2 through 7×7, Megaminx, Square-1.
The three-body problem — chasing conserved quantities in deterministic chaos.
Sums of sines — turning prime numbers into music, and signal out of noise.
Foundations
Machine Learning & CS
Chemistry
Biology & Physics
Shortcuts & secrets