Hi, I'm Deven. I work on AI safety, mostly interpretability, with a growing focus on scheming, evals and control.
I test claims about AI models against answers I already know: a motive I planted, a surname I swapped, an explanation I can check against the model’s own features.
What I've found so far
Quiet Motive
I gave a model a hidden motive and asked another model to work out why it misbehaved. Showing the investigator the model's reasoning changed nothing, and the published method's counterfactual edits flipped 52 of 87 verdicts for a net of zero.
MATS 12.0 work test, top 6.1% of 888 applicants · Blog post · Code
Loyal Lies
With Adhiraj Pundir. We fine-tuned Qwen-2.5 models with hidden objectives and tried to catch them with probes. The probes picked up how visible the behaviour was rather than the intent behind it.
5th of 179 teams, Apart Research Secret Loyalties Hackathon · Write-up · Code
Crosswise
A pip-installable tool that checks natural-language explanations of activations against SAE evidence, to catch explanations that sound right and aren't.
83% recall at 67% precision on a 212-position benchmark · Blog post · Code
AuditAI
A caste-bias audit of Gemma-2-2B using byte-identical counterfactual surname pairs.
Top 100, Google Solution Challenge 2026 · Code
Right now
October 2026
- On 7–8 November I'm giving a five-minute talk at Swiss AI Safety Days in Zurich: "It Found the Evidence, Then Ignored It".
- Saturdays go to TARA's Delhi cohort, working through the ARENA curriculum until December.
- My internship is at Authentic Memory, a Munich nonprofit building cryptographic provenance infrastructure for Holocaust testimony and historical archives.
- I'm planning a 2–3 month interpretability project aimed at ICML 2027 and looking for one or two collaborators. If you have an idea you'd like to test properly, email me.