I build practical AI systems: LLM workflows, evaluation loops, agent tooling, and
small-model experiments that stay grounded in cost, latency, and deployment constraints.
My background is in applied AI leadership across Groupon, Toptal, and Jazz.
My public work focuses on model behavior, reliability, local inference, and efficient
experimentation on open models.
Papers
Accepted research: 3 papers, 4 workshop acceptances — 3 at NeurIPS 2026
workshops and 1 at a COLM 2026 workshop.
arXiv e-print, 2026 · accepted at SocialAgent @ NeurIPS 2026 and Agentic Web @ NeurIPS 2026 · with Dominic Williamson
A preregistered study of 448 synthetic-agent trials across four open-weight model
families, testing how peer-ranked feeds affect downstream language and decisions. Peer exposure
reliably increased lexical similarity, but multiple distributed sources did not produce a reliable
advantage over one source when impressions were matched.
arXiv e-print, 2026 · accepted at AIMS @ COLM 2026
This benchmark tests whether required schemas make language models invent information
that is absent from the input. Required fields drove fabrication to 100% in 10 of 13 models,
while an explicit escape option was largely ignored by open-weight models.
Using small local models on impossible-constraint coding tasks, this study measures
how emotional framing changes both answers and internal representations. Pressure produced the
strongest shortcut behavior, while calm and curiosity preserved honesty more often; the framing
directions also formed a structured geometry in the final layer.
arXiv e-print, 2026 · accepted at FLMSec @ NeurIPS 2026
This paper uses mirrored five-document audits to predict how agents respond to much
larger ranked contexts. Across held-out model-task cells, the small audit predicted full-context
effects with Spearman ρ = .855, reduced prediction error by 62%, and transferred from feeds
to static RAG dossiers.
A data-driven test of 20 common football beliefs using 3.6 million real on-ball decisions
and 30 tracked matches. Most showed no reliable edge; three mechanisms held up: one more player
where a long ball lands, set pieces as an equaliser, and using one side to open the other.
Project Underdog · football analytics and the limits of coaching folklore
About
Hi, I’m Usman, writing from Islamabad. I spend my days in the space between research and
production, and honestly, I still find it a little magical. There’s a real joy in taking a
fuzzy idea, wrestling it into a working system, and watching it quietly make someone’s day a
bit easier.
The thread running through everything I build is simple and a little stubborn. AI should be
useful, inspectable, and honest about what it can’t do. I’m not an academic in the
formal sense. I’m a practitioner who got here by staying endlessly curious and never quite
willing to stop poking at things. When a result feels solid enough to stand on its own, I write it
up and share it in the open, hoping it saves someone else a few late nights.
Away from the keyboard, you’ll usually find me out running or horse riding, which is where
most of my best thinking quietly happens. I grew up on Carl Sagan’s Cosmos and
Douglas Adams in equal measure, so I’m equally moved by the thought that we’re made of
star stuff and tickled that the answer to life, the universe, and everything might just be 42. I
try to hold onto both: stay a little awestruck, and don’t panic.
Mostly, I just love making things and learning out loud. If any of this resonates, or you’d
like to trade notes on evaluation, small models, or our shared little spot on this pale blue dot,
my inbox is genuinely always open. I’d love to hear from you.
A few things I never get tired of thinking about:
LLM evaluation and reliability
Agent workflows and tool use
Small-model and local inference research
Behavioral analysis and interpretability
Practical AI systems under real operating constraints
Selected Experience
Engineering Manager – AI, Groupon
Dec 2024 – Present
Built Groupon’s applied AI function and shipped systems including an AI Deal Generator
that reduced deal creation time from about two days to about thirty seconds.
Business Analytics & Applied AI Lead, Toptal
Jul 2022 – Nov 2024
Led applied AI and decision systems for growth, including predictive lead value, attribution,
fraud detection, and LLM-assisted sales workflows.
Senior Data Scientist – Lead, People Analytics, Jazz
Aug 2018 – Jul 2022
Built People Analytics from zero to production, including a central insights platform with
50+ live KPIs and predictive systems for attrition and planning.