Charles N. Christensen

I’m a research scientist at GSK in London, working on AI agents for long-horizon scientific research. My focus is agent memory, autonomous research loops, self-improving harnesses, and continual learning from an agent’s own experience, without weight updates.

I lead GSK’s agentic capabilities research team and I’m one of the leads on Cogito Forge, GSK’s AI scientist. I have worked on every component of it since its first week, and its memory is the part I research. My PhD at Cambridge was in machine learning for super-resolution microscopy.

Charles N. Christensen
  • Agent memory
  • Autonomous research
  • Self-improving harnesses
  • Learned orchestration
  • Continual learning
  • RL post-training

Research

I lead the agentic capabilities research team at GSK. We work on agents that run experiments on their own and agents that keep improving after they ship.

Autonomous research

These agents run the research loop themselves. They propose an experiment, run it on the GPU cluster, read the result and choose the next one. One of ours reproduced a published RL result claim by claim, then extended it across a tree of budgeted experiments.

Self-improving harnesses

A frontier model sits in an outer loop and redesigns the agent organisation between evaluation rounds. We compare its designs with hand-built, one-shot and RL-trained ones on held-out life-science tasks.

Learned orchestration

We train a planner with RL to design the agent graph for each task. It chooses between DAGs, trees and cycles, and a model-agnostic runtime executes the design.

GRPO, REINFORCE, LoRA

Continual learning

Agents can improve with use without any weight updates. We work on learned consolidation and forgetting across sessions, a skill library distilled from successes and failures, and a learned controller that decides what to keep under a fixed memory budget.

Post-training and RL

We post-train LLMs with reinforcement learning and preference optimisation on a Slurm GPU cluster. The work covers policy-gradient methods, reward models and constrained PPO.

RLOO, DPO, PPO

Inference and tool use

Before the AI scientist, I engineered multi-node inference for GSK's in-house LLMs, with constrained decoding for tool calls, and built research agents for Jules, GSK's agent platform.

Open source

I build my own tools and publish most of them. Some are research code with papers behind them. Others exist because I wanted them on my own machine.

Research code

ML-SIM

ML-SIM reconstructs super-resolution microscopy images with a model trained only on simulated data. It has never seen a real microscope image, and it reconstructs real data in under 200 ms. Published in Biomedical Optics Express and cited more than 90 times.

ERnet

ERnet segments the endoplasmic reticulum in live-cell video with a vision transformer, then turns the network into a graph so its remodelling can be measured. The graph extraction is tested on synthetic networks with a known graph, generated as shown. Published in Nature Methods.

VSR-SIM

Structured illumination microscopy needs nine exposures of a sample that holds still, and living cells do not. VSR-SIM is a video transformer that reconstructs through the motion and returns an image for every raw frame instead of one for every nine.

Tools and side projects

LLMux

Ask ChatGPT, Claude and Gemini the same question side by side in one window, or switch to just the one you need.

coding-agent-tools

Search and resume sessions across Claude Code, Codex and OpenCode from one command, with token and cost analytics for all three.

STT latency benchmark

I timed 6 speech-to-text APIs on the same audio from the same machine, back to back, to find the fastest for real-time use. The chart shows median end-to-end latency.

embedded-3d-rendering

A 3D graphics pipeline in C for a microcontroller with 2 kB of RAM. It has a scanline rasteriser and a depth test for every pixel, with no GPU and no frame buffer.

Minimal Deep Research

A small implementation of Deep Research. It plans queries, searches, summarises, finds the gaps and goes again before it writes the report. Works with any OpenAI-compatible model.

LangGraph code agent

A code agent that breaks a programming task into steps, then generates, runs and validates Python in a REPL until the task passes.

The rest is on GitHub

Publications

Before agents, I worked on deep learning for microscopy. I built reconstruction, denoising and segmentation methods that hold up at low signal-to-noise, trained on physically modelled synthetic data.

  1. Nature Communications, 2022

    Machine learning assisted interferometric structured illumination microscopy for dynamic biological imaging

    Edward N. Ward, Lisa Hecker, Charles N. Christensen, et al.

  2. arXiv preprint, 2022

    Spatio-temporal Vision Transformer for Super-resolution Microscopy

    Charles N. Christensen, Meng Lu, Edward N. Ward, Pietro Lio, Clemens F. Kaminski

  3. Biomedical Optics Express, 2021

    ML-SIM: Universal reconstruction of structured illumination microscopy images using transfer learning

    Charles N. Christensen, Edward N. Ward, Meng Lu, Pietro Lio, Clemens F. Kaminski

  4. Science Advances, 2020

    The structure and global distribution of the endoplasmic reticulum network are actively regulated by lysosomes

    Meng Lu, Francesca W. van Tartwijk, Julie Qiaojin Lin, … Charles N. Christensen, et al.

  5. Applied Optics, 2019

    Simple and robust speckle detection method for fire and heat detection in harsh environments

    Charles N. Christensen, Yevgen Zainchkovskyy, Salvador Barrera-Figueroa, et al.

  6. Physical Review A, 2018

    Driving-induced population trapping and linewidth narrowing via the quantum Zeno effect

    Charles N. Christensen, Jake Iles-Smith, Torkil S. Petersen, Jesper Mørk, Dara P. S. McCutcheon

In preparation

  • Learned memory control for LLM agents under a hard budget.
  • Cogito Forge, GSK's AI scientist, and its evaluation.

Full list on Google Scholar

Physics, microscopes, then agents

  1. 2023 to now

    Senior AI/ML Research Scientist

    GSK, London

    I have been one of the leads on Cogito Forge since its first week, and since 2025 I have run the agentic capabilities research team. Before that I worked on LLM inference and research agents for GSK's Jules platform.

  2. 2022 to 2023

    Machine Learning Engineer

    Oxford Nanoimaging (ONI), Oxford

    I built deep learning for spectral demixing, segmentation and super-resolution in new microscopy products, and took it from research to deployment.

  3. 2019 to 2022

    Co-founder

    Wizion AI, Cambridge

    A colleague and I commercialised the deep-learning image processing from our lab. We built 2 products and took part in the Cambridge Judge Business School Accelerate programme.

  4. 2018 to 2022

    PhD

    University of Cambridge

    My thesis was on deep learning for super-resolution microscopy, supervised by Clemens Kaminski and Pietro Liò. The methods are released as ML-SIM, VSR-SIM and ERnet.

  5. 2018 to now

    Co-founder

    open-seneca

    We build low-cost air quality sensors and run citizen-science networks with them in cities on 4 continents. It started as a research project during my master's degree.

  6. 2012 to 2018

    Physics, then mathematical modelling

    Technical University of Denmark and Cambridge

    A BScEng in physics, an MScEng in mathematical modelling and computation, then an MRes at Cambridge. I worked on quantum optics, fibre lasers and nanomagnetism before I got to machine learning.

Contact

If you work on agent memory, autonomous research or agent evaluation, I'd like to compare notes. Email is the fastest way to reach me.

[email protected]