Work

Checkpoint-based LLM router that separates labelled data, training and runtime model selection. I co-founded Nordlys Labs and led its AI routing work through May 2026.

Audits LLM judges for position, length, self-preference and consistency failures. On 150 JudgeBench pairs, Ministral-8B fell from 0.61 raw accuracy to 0.12 correct in both answer orders.

Benchmarks KV-cache compression strategies and their quality trade-offs. Stratigraphic reported a 2.2× smaller analytical KV cache for +0.04 perplexity on Llama-3.2-1B-Instruct.

Repository-analysis agent that explores code and git history incrementally through a REPL. It provides 12 git-aware tools, sub-model delegation, trajectory logs and hard cost budgets.

Internal-state fuzzing framework for LLM inference engines. It snapshots and mutates KV-cache, scheduler and prefix-cache state, then checks invariants through reusable oracles.

Synthetic month-end reconciliation benchmark with a deterministic ledger verifier, tool-using agent harness and hard gates for unauthorised accounting changes.

Simulation study of football and backflip behaviours for Open Duck Mini. A trained keep-up policy averaged 4.87 consecutive touches; analysis showed why the current robot cannot backflip.

Compares how Python libraries are documented with how they are used in public repositories. A reference run analysed 119 repositories and 544 call sites.

Browser-based 3D CSV explorer with on-device hand tracking. It renders a documented 2,200-point dataset in one instanced draw call across three graph topologies.

Voice-assisted medical-sales prototype combining push-to-talk interaction, transcription and PubMed literature lookup.

Simulation prototype exploring adversarial sensor inputs and reinforcement-learning task planning for autonomous satellites.

From-scratch learning implementation covering probability, KL divergence, ELBO, VAEs and DDPM image generation.

Experimental cross-domain idea generator that uses structural analogies, critique and falsifiable predictions.

Small-model replication of prompt repetition with paired McNemar tests and explicit negative results.

Prototype for an advertising-supported AI chat experience with contextual ad placement.

Reusable agent workflow for business discovery, research, option comparison and implementation planning.

Lead-discovery prototype for pharmaceutical sales, combining map search with LLM-assisted query refinement.

University research project on ML-based intrusion detection for in-vehicle networks.

From-scratch Transformer implementation covering attention, positional encoding and training.

From-scratch LSTM implementation covering gates, cell state and backpropagation through time.

Mini3ADR lexer and parser built with Flex and Bison for a compiler-construction course.

Character-level RNN text generator used to study sequence modelling and vanishing gradients.