Skip to the page
A. Flores Sepúlveda
Research notebook · Page 1Bern · 2026

Alejandro Flores Sepúlveda1,2,3,†

  • 1Reymond group, University of Bern. PhD candidate, 2024–2028; currently on generative methods.
  • 2SPAR. Research fellow in AI safety, Fall 2026; escalation detection in AI agents.
  • 3Mafer.ai. AI Engineer from October 2026; food & fragrance, across product, engineering and research.
  • †1st on the Prosper AI track, HackSpain, Madrid, September 2026; grand finalist (5 of ~60 teams).

AbstractI’m a technical generalist working across machine learning for drug discovery, AI safety and generative methods, and I build products at hackathons. Chemistry and large chemical maps are where I started; they are one thread among several. This site collects the projects, the papers and the long-form notes. Figure 1 is a game; play as long as you like.

Keywords:
Score 0Best 0Rings 0
Chain CCC
Figure 1. SMILES Snake. Eat atoms to grow a chain, and bite your own body to close a ring. On a square grid every loop has an even number of atoms, so rings come in fours, sixes and eights; six is the most stable. Click the figure, then steer with the arrow keys or WASD. P pauses.
Ring catalogue 0 of 26 found

Every named ring you close stays here, on this device. Unfound rings show only their size and heteroatoms.

    1Now

    • from Oct 2026AI Engineer at Mafer.aiFood & fragrance: product, engineering and research.
    • Fall 2026SPAR research fellow, AI safetyEscalation detection in AI agents.
    • 2024–2028PhD candidate, Reymond group, University of BernCurrently working on generative methods.
    • Sep 2026HackSpain, Madrid1st on the Prosper AI track; grand finalist, 5 of ~60 teams.

    2WorkProjects and research, newest threads first. The project pages.

    • Escalation detection

      Can turn-averaged sparse autoencoders tell when an AI agent starts taking more extreme measures to get past an obstacle, and when it should pause and hand the problem to a human?

      AI safetySPAR · Fall 2026
      in progress
    • ConceptSteer

      Model diffing and activation steering on a chemical language model, with a sparse autoencoder trained from scratch.

      AI safety · Drug discoveryin progress
    • Generative methods

      Current PhD work in the Reymond group.

      Drug discoveryPhD · 2024–2028
      in progress
    • Chelombus

      A searchable map of 9.6 billion drug-like molecules from Enamine REAL. Runs end to end in about 4.5 hours on one desktop GPU; it used to take about 14 days.

      Drug discoveryJ. Chem. Inf. Model. 2026
      chelombus.gdb.tools
      180k+ maps served
    • TMAP 2.0

      A rewrite of the TMAP layout library in Numba, about 3× faster. Maps for datasets too large to read as tables.

      Drug discoverymanuscript in preparation
    • SmilesDrawer

      Turns SMILES strings into molecule drawings in the browser. Ring perception is now rebuilt on Vismara’s algorithm.

      Drug discovery~26k npm downloads a week
    • HackSpain 2026 project

      1st on the Prosper AI track and a grand finalist. Write-up coming.

      BuildingMadrid · Sep 2026
    • REFUGIO

      A hackathon project, with a postmortem built from interactive widgets.

      Buildingpostmortem · Jul 2026

    3WritingInteractive notes, newest first.

    4Papers

    1. Alejandro Flores, Jean-Louis Reymond. Chelombus: Clustering and Nested TMAP Visualization of Billion-Scale Chemical Spaces. J. Chem. Inf. Model. 2026.doi:10.1021/acs.jcim.6c00420 ↗
    2. Aline Carrel, Alejandro Flores, Maedeh Darsaraee, Jean-Louis Reymond. Spirocyclic Diamine Scaffolds for Medicinal Chemistry. Helv. Chim. Acta 2025.doi:10.1002/hlca.202500207 ↗
    3. Luis Apaza Ticona, Javier Sánchez Sánchez-Corral, Alejandro Flores Sepúlveda, Carmen Soriano Vázquez, Carmen Hernán Vieco, Ángel Rumbero Sánchez. Novel 1,2,4-oxadiazole compounds as PPAR-α ligand agonists: a new strategy for the design of antitumour compounds. RSC Med. Chem. 2023, 14, 1377–1388.doi:10.1039/D3MD00063J ↗

    5Contact

    Write to afloresep01@gmail.com