· Conference notes
What I took away from RecSys 2026
Personal reflections on the modeling ideas and practical optimizations that stood out.
Read the articleMachine learning · Research · Systems
Ghulam Ahmed
Ansari.
Building systems that
learn from feedback.
I'm a Staff Machine Learning Engineer at LinkedIn. I work on recommendation, exploration, and evaluating what works, with research roots in reinforcement learning and neuro-symbolic reasoning.
Ghulam Ahmed Ansari Sunnyvale, California
Research meets practice
LinkedIn 2020–present
IBM Research 2017–2019
IIT Madras B.Tech + M.Tech
Writing
01 / About
From a research idea
to a working system.
My work sits between machine learning research and the engineering it takes to put that research into practice.
At LinkedIn, I've worked across Feed ranking, content exploration, off-policy evaluation, Learning Search, and misinformation detection. That has meant both building models and building the infrastructure to understand how well they work.
Before LinkedIn, I worked at IBM Research on learning to answer complex questions by generating executable programs. Earlier, at IIT Madras, I studied how reinforcement learning agents could share knowledge across text-based games.
The settings are different, but a recurring question connects them: what can a system learn from the feedback it has, and how can we tell whether it is making better decisions?
02 / Selected research
Ideas, with the papers behind them.
Interference through Shared Model State: A Shift-Share Design
Evaluating recommendations at scale
From M Passes to One
A MapReduce-Style Bootstrap for Pointwise Off-Policy Evaluation in Production Recommenders
Estimating a recommendation policy's performance from existing logs is only part of the problem. This work makes estimating the uncertainty around that result more practical, computing multiple bootstrap estimates in one pass over the data.
Read the paper Off-policy evaluation · Scalable estimation
Learning when feedback is noisy
Neural Program Induction for KBQA Without Gold Programs or Query Annotations
Learning to generate programs for question answering becomes harder when input annotations are noisy. SSRP uses continual retrospection, comparing current learning with a reference model, to reduce noise-induced instability.
Read the paper Program induction · Sparse rewards
Reasoning through executable programs
Complex Program Induction for Querying Knowledge Bases in the Absence of Gold Programs
How can a model learn multi-step reasoning without examples of the correct programs? CIPITR combines neural program generation with symbolic constraints and auxiliary rewards to search for executable answers.
Read the paper Weak supervision · Knowledge bases
Sharing knowledge across environments
Language Expansion In Text-Based Games
Can one agent learn from several games with different vocabularies? This work studies policy distillation from specialized agents into a shared agent, using text-based games to examine what transfers.
Read the preprint Policy distillation · Language
03 / The journey
Different problems.
A connected path.
From electrical engineering in Chennai to research in Bengaluru and recommendation systems in California.
Full experience on LinkedIn2024 – presentStaff ML Engineer LinkedIn · Sunnyvale
Technical leadership in Feed ranking and content discovery, connecting machine learning, infrastructure, product, and data science.
My work spans the models that select content and the systems needed to evaluate and serve them.
2022 – 2024Senior ML Engineer LinkedIn · Sunnyvale
Built off-policy evaluation and exploration systems for recommendation. This brought together statistical estimation, calibration, contextual decision-making, and large-scale data processing.
2020 – 2022Applied Research & AI Engineering LinkedIn · Bengaluru
Worked on LinkedIn Learning Search, then helped build the Misinformation AI team. Projects included search ranking, offline replay evaluation, image near-duplicate detection, and graph-based detection methods.
2017 – 2019Research Engineer IBM Research · Bengaluru
Studied reinforcement learning and neuro-symbolic methods for knowledge-base question answering. This work led to CIPITR and SSRP, published in TACL and IJCAI.
2012 – 2017B.Tech + M.Tech Electrical Engineering · IIT Madras
Moved from electrical engineering into machine learning and reinforcement learning. My master's research with Balaraman Ravindran explored policy distillation and language expansion in text-based games.
04 / Beyond the papers
Invention and community.
Selected granted patents
Co-inventor on work in visual reasoning and interactive image generation.
US 12,367,242 · Granted 2025Decentralized online multi-agent visual question answering US 11,335,084 · Granted 2022Image object anomaly detection US 11,074,735 · Granted 2021Multistep interactive image generation utilizing knowledge storeRecognition
KDD 2025 Excellent Reviewer
Recognized for reviewing in the Applied Data Science Track.
Conference listingProfessional community
IEEE Senior Member
Alongside engineering and research, I contribute through peer review. My reviewing has included KDD, ACL's Student Research Workshop, ICLR workshops, and CHIL.
Volleyball every week.
I also enjoy football, trekking, scuba diving, photography, and a game of Catan.
05 / Connect
Let's compare notes.
Recommendation, reinforcement learning, evaluation, or the engineering between them.
Connect on LinkedInA bio to take with you
For talks & introductionsGhulam Ahmed Ansari is a Staff Machine Learning Engineer at LinkedIn, working on recommendation, exploration, and off-policy evaluation. Previously at IBM Research, he studied neuro-symbolic reasoning and reinforcement learning. He holds B.Tech and M.Tech degrees in Electrical Engineering from IIT Madras.