The bookshelf
One paper out, five under review. I write down my predictions before running the experiments that test them, and I report the ones that fail.
- Published, Chemical Engineering Science 2025
Affordable Smartphone-Assisted Diagnostics: Computer Vision on Paper Microfluidics for Uric Acid Detection
Uric acid is a key biomarker: too high points to gout, hypertension and kidney disease, too low to neurodegenerative disorders. This work pairs a paper microfluidic device with a smartphone to measure uric acid between 1.5 and 25 mg/dL, with no external light source, enzymes or nanoparticles. The channels are drawn onto filter paper by a low-cost DIY XY plotter using technical pens filled with a hydrophobic PDMS solution. A chromogenic reaction turns the sample bluish-green, and an Android app reads its luminance against eight on-chip reference zones. Built with NPL India and National United University, Taiwan; two patents filed.
Read the paper - Under review, ICLR 2027
Looped Transformers Learn One of Two Algorithms, and Training Moves the Odds
Depth-recurrent (looped) transformers trained on prefix-product tasks do not learn a spectrum of more or less efficient solutions. Causal interventions show that each model installs one of two discrete algorithms, and the bimodal split replicates across 237 and 90 independently measured checkpoints. Interventions on capability do not move the odds between the two; only making one of them infeasible does. Every claim was registered before the run that tested it, and all fourteen failed predictions are scored in the paper.
Ask me for the draft - Under review, TMLR 2026
Sign Conflicts in LoRA Adapter Merging Are Pervasive, Not Localized
Sparsification-based merging methods assume that sign conflicts between LoRA adapters are concentrated in a few weights that can be trimmed away. Across five LLMs they are not: about 73% of weights conflict, close to the random-sign baseline, and the pattern follows the training seed rather than the task. The verdict holds on a deliberately divergent math, code and sentiment arm, and across sweeps over adapter rank and LoRA scaling.
Ask me for the draft - Under review, ARR for EACL 2027
Weight-Space and Behavioral Interference in Multi-Objective LoRA Debiasing for Indian Social Bias: A Cross-Architecture Study
MOOLoRA treats debiasing Indic-language models across four Indian social axes as multi-objective optimisation over LoRA merge coefficients, solved with NSGA-II. Across four backbones, the weight-space alignment between per-axis adapters looks the same everywhere, but behavioural spillover does not: it can reverse sign between architectures whose weight alignment is statistically indistinguishable, and no weight-space quantity tested predicts its direction. The paper also identifies a coherence failure in naive Pareto evaluation and redesigns the protocol.
Ask me for the draft - Under review, IEEE TSE 2026
Same Models, Different Verdicts: An Evaluation-Sensitivity Study of LoRA Adapter Merging for Multi-Task Code Analysis
An audit of 1,020 configurations of LoRA adapter merging for code-analysis tasks. Three of five common evaluation-pipeline perturbations silently reverse the top-line conclusion about which merging method is best, so the choice of evaluation pipeline can matter as much as the choice of method.
Ask me for the draft - Preprint, Zenodo 2026
A Rate-Distortion Function for Model Merging
Practitioners merge many LoRA fine-tunes into one deployable model, but no principled limit says how well this can be done at a fixed storage budget. This paper casts merging as multi-source lossy source coding under worst-task distortion and proves the first rate-distortion theorem for it: a closed-form floor set by the geometry of the task subspaces, plus a compression term that shrinks exponentially with the bit budget. A random rotation followed by uniform scalar quantization matches the lower bound up to a constant. Across 16 real LoRA adapters the floor is zero, yet current methods still leave 0.10 to 0.22 nats per token of worst-task error, so the gap is algorithmic, not informational.
Read the paper