Dakshesh Dharmadhikary
Summary
I'm the co-founder and CEO of Radiant, where we build infrastructure to run, observe and improve AI agents in production, with a focus on voice and long-running agents in regulated settings such as healthcare and finance. My research looks at how agents fail over long interactions: how models give in under sustained pressure, what that looks like inside the model, and whether steering or learned world models can fix it. Before that I studied physics at Oxford, where my MPhys research was on perovskite solar cells.
Research
- Single-reply calibration underestimates the multi-turn cost of activation steering November 2026 • Research post, Radiant • dakshe.sh/steering
- Looking inside an agent as its boundaries erode October 2026 • Research post, Radiant • dakshe.sh/boundaries
- Stress-testing AI agent emotional boundaries: smaller models outperform March 2026 • Radiant Labs Substack • dakshe.sh/stress-test
- Wide band gap hybrid lead halide perovskites for tandem solar cells 2022 • MPhys thesis, University of Oxford • supervised by Benjamin Putland, Dr Harry Sansom and Prof Henry Snaith • dakshe.sh/wide-bandgap-perovskites-thesis
Experience
Radiant (formerly ClimaView): Co-founder & CEO
London, UK | Jan 2023 – Present · CEO since Nov 2025; previously Chief Product Officer
- Lead a three-person applied AI research team; raised venture and angel funding from investors including Plug and Play, Founders Factory and SoftBank.
- Built Radiant Platform (agent execution, tracing, evaluation and a Meta Agent for failure analysis) and the open-source Polaris agent framework.
- Shipped production agents: referral-booking voice agents at a US hospital group (85% less patient admin), fine-tuned LLMs for transaction analysis at a UK bank (90% lower inference cost), and compliance and M&A research agents (80% less work, 10× faster).
Aurora Energy Research: Data Science Intern
Oxford, UK | Jul – Sep 2021
- Improved wind-energy forecasts through statistical downscaling; built Python/SQL pipelines for geospatial processing and evaluated predictions against wind-speed observations.
Cranfield University: Research Intern, Photonics
Jul – Aug 2017
- Developed LabVIEW control and acquisition software for a Mach–Zehnder optical-fibre interferometer measuring displacement at nanometre resolution.
Transport Systems Catapult: Software Engineer Intern
Jul – Aug 2016
- Built a Unity3D/C# and Arduino VR cycling simulation to study interactions with autonomous vehicles; prepared datasets for machine learning.
Education
University of Oxford, Keble College: MPhys Physics
Oxford, UK | 2018 – 2022
- Gibbs Prize for Outstanding Academic Achievement; S. S. Chang Memorial Prize for Entrepreneurship.
- MPhys research on band-gap tuning in mixed-halide perovskites for tandem solar cells (thesis): fabricated thin films and single-junction devices, characterised them with UV-vis, photoluminescence, XRD and electron microscopy, and reached 9.53% power conversion efficiency.
The Royal Latin School, Buckingham
2011 – 2018
- A-Levels: A*A*A*A (Physics, Mathematics, Further Mathematics, Computer Science); 11 A* GCSEs.
Skills
- AI systems: agent architectures • LLM evaluation • interpretability and activation steering • fine-tuning • retrieval • model deployment
- Programming: Python • Rust • TypeScript/Svelte • SQL