LLM research & engineering
Co-founder of enrik.ai, an AI-native go-to-market (GTM) platform.
My research interests are LLM training, evaluation, and reliable reasoning. My public work includes language-model experiments, AI agent security, and tools for coding agents.
I won a gold medal and placed 9th out of 4,211 in the Multi-Step Tool Attacks competition. I collaborated with Refractal AI, a London-based AI security startup, on the competition.
The public solution includes a technical working note and evaluation tooling for the submitted attack, reference defenses, and candidate search. The write-up documents the agent-assisted workflow, attribution, and limits of local replay.
Independent experiment · July 2026
I evaluated Qwen2.5-7B-Instruct on a bookkeeping task with reasoning depths of 4, 8, 16, and 32, using 32 tasks per condition and depth. The experiment compares chain-of-thought prompting with an external state ledger and ablations of its checking mechanism.
At depth 32, the ledger condition reached 50.0% accuracy versus 28.1% for the chain-of-thought baseline. Removing the checking mechanism produced the same accuracy at every tested depth: the checks added no measured accuracy in Stage A.
These preliminary results cover a single model and task family. This is an inference-time evaluation using a fixed pretrained model. The report includes the experimental setup, confidence intervals, ablations, and limitations; code and recorded results are public.