Research.

Formal work moves slowly and becomes visible only at the end. This page keeps score honestly: what's accepted, what's done, and the questions that stay open longer than any paper.

Accepted

Two papers accepted in 2026. One with IEEE, on scheduling LLM inference without ever reading the prompt. One with Springer, a pre-registered study of what translation does and does not fix when populating Wikidata from eight Indic languages.

read → the Springer paper · the IEEE paper

Recent work

Decomposing the Low-Resource Penalty in Indic Knowledge Graph Population

multilingual NLP · accepted

A pre-registered study of what translation does and does not fix when populating Wikidata from eight Indic languages.

Quantifying the Cost of Content-Blindness in LLM Inference Scheduling

cloud scheduling · accepted

What LLM inference scheduling loses when it cannot read the prompt, measured on 44.1M real Azure production requests. It is less than the literature assumed.

Clinical risk prediction

applied ML · complete

Pressure-ulcer risk modeling with ensemble methods — audited end to end and rebuilt where the audit demanded it.

Provenance for code

protocol design · in progress

The research half of Etch: what a checkable claim of authorship requires, and how small the trusted surface can get.

Full publication list available on request.

Open questions

q01

who owns a refactor?

Git credits the last hand to touch a line — so a refactorer quietly inherits authorship of code they never designed. Any provenance system that can't tell restructuring from authorship is measuring the wrong thing.

q02

does content-blind scheduling survive prefix caching?

Multi-turn conversation traffic serves repeated prefixes from cache, and the traces used to measure content-blind scheduling record context length but not cache residency. Model a 90% hit rate and the measured improvement over first-come-first-serve falls hard. This is the open problem that work leaves standing, not a footnote to it.

q03

how do you name an entity you have only seen in another script?

Translating the input repairs relation extraction and does nothing for entity grounding, because understanding a sentence and canonicalizing the things in it are different skills. A model that reads a Tamil sentence perfectly can still fail to produce the right Wikidata identifier. Canonicalization across scripts is the half of the problem no amount of translation reaches.

q04

what does provenance mean when the author is a pipeline?

When a model writes the code and a human approves it, neither is cleanly the author. Provenance may have to record a chain of custody — who signed off, under what key — rather than a single name.

q05

how much of infrastructure is secretly a pricing problem?

More than anyone admits. Rate limits, quotas, priority classes, retry budgets — each is a price refusing to call itself one. See contention as a market and half of infrastructure design turns out to be economics in a trench coat.

linger on a question for where the thinking stands — they outlive any answer, though; a good question is a research program wearing modest clothes