Latest · October 2, 2026

Publications

Most Recent

10/02/26CODE—

DeferSeg

Tile-scheduled neural passes for real-time compositing.

09/23/26PAPER—

Notation matters: cross-representation inconsistency in chemistry language models and its mechanistic origins

Digital Discovery, Royal Society of Chemistry (Gold Open Access). The same molecule written as SMILES, IUPAC, InChI or SELFIES gave inconsistent predictions for 88% of 1,072 molecules, and ChemBERTa-2 representations of identical molecules diverge layer by layer.

09/20/26CODE—

Fruit-fly connectome trader

A weekend project with Krishang Sharma: the fruit-fly mushroom-body circuit wired into a trading system on Binance Spot Testnet, with dopamine-gated plasticity driven by realised P&L. Open-sourced on GitHub.

09/04/26PAPER—

Translation Metrics Cannot Judge What Their Tokeniser Deletes

WMT 2026 (poster), co-located with EMNLP 2026, Budapest. LIGATUR and AEGIS submissions to the shared task on automated translation quality evaluation, with Krishang Sharma and Sonia Khetarpaul.

09/04/26DATASET171 items

LIGATUR

A contrastive challenge set (English-German, English-Hindi) separating Unicode-level corruption from semantic error, released with the Palimpsest code and the AEGIS metric.

08/22/26PAPER—

Beyond Epistemic Collapse: Disagreement-Aware Scientific Retrieval-Augmented Generation

EMNLP 2026, Main Conference (CORE A*). EVIRAG keeps conflicting scientific evidence apart instead of collapsing it into one answer, and improves contradiction recall and viewpoint coverage over standard RAG. With Krishang Sharma and Sonia Khetarpaul.

08/22/26DATASET1,250 queries

EVIRAG-BENCH

A benchmark spanning five scientific domains, with a Rust pipeline, evaluation metrics and corpus tools.

07/26/26CODE—

Machina (ClerkTree)

ClerkTree's first public research release: open machine intelligence for industrial machines, covering bearing-fault classification, remaining useful life, visual quality inspection and evidence-grounded industrial reasoning.

05/24/26AGENT—

Juris on IBM VAKRA

Ranked #2 globally for tool selection. A 36B model with a capability-specific routing stack over roughly 8,000 APIs in 62 domains: shortlist the tools, force one choice, normalise arguments, return answers directly from tool output.