Close

Ghulam Ahmed Ansari

Staff Software Engineer, Machine Learning

Download Resume   GitHub   LinkedIn  OpenReview Google Scholar Profile  

About Me

MyPic

I am a Staff Software Engineer - AI at LinkedIn, where I lead teams focused on improving topical relevance and user engagement. My work on the Topicality L1 initiative has driven measurable impact, including a +0.09% increase in Feed Sessions and a +74.13% Topicality Viral Action Rate. Previously, as a Senior Software Engineer, I designed and implemented a scalable Off-Policy Evaluation (OPE) framework, improving offline-online metric consistency by over 600% and led the Gold Content Explore-Exploit initiative, increasing Daily Unique Gold Professional Interactions by 5.6%.

Before LinkedIn, I was part of the founding Misinformation Detection Team at LinkedIn Bangalore, developing AI-driven solutions for early misinformation detection. Prior to that, I worked as a Research Scientist at IBM Research Labs, focusing on deep reinforcement learning for knowledge-based question answering (KBQA). My research spans Neuro-Symbolic Methods, Multi-Agent Learning, and Large-Scale Recommender Systems, with publications in NeurIPS, IJCAI, ACL, and TACL.

I am passionate about bridging research and industry to build impactful AI systems, particularly in recommender systems, off-policy evaluation, deep reinforcement learning, and scalable machine learning infrastructure. I actively contribute to the research community through peer reviews for top-tier conferences and am always open to discussions on advancing AI-driven personalization and decision-making systems.

Experience

LinkedIn, Sunnyvale, CA

Staff Software Engineer - Machine Learning

  • Leading 10+ engineers across multiple teams as part of the Topicality L1 (First Pass Ranker) Working Group for LinkedIn’s most ambitious AI-first session-targeting initiative aimed at accelerating topical relevance across multiple surfaces to drive Feed Sessions and Weekly Active Users.
  • Achieved notable outcomes from combined MME initiatives: +0.09% Feed Sessions, +74.13% Topicality Viral Action Rate, and +20.37% Topicality Feed Updates Viewed over three quarters.
  • Quickly ramped up on the Topicality L1 stack, scoped initiatives, and negotiated with Product to finalize initiatives.
  • Collaborated closely with AI, infra, and data science stakeholders to ensure accountability, unblock teams, and communicate delays/blockers to leadership.

LinkedIn, Sunnyvale, CA

Senior Software Engineer - Machine Learning

  • Designed and implemented the Off-Policy Evaluation (OPE) Framework, boosting offline-online metric consistency by +600% over legacy methods for main feed evaluations.
  • Developed a robust model-free calibration technique that supplements OPE, providing accurate estimates despite significant shifts in underlying architectures and multi-objective weights across models.
  • Built a scalable Bootstrap Inverse Propensity Scoring algorithm to process billions of historical logs under varied hyperparameters, enabling robust performance estimates and confidence bounds.
  • Filed two patents related to OPE and multi-objective weight optimization, significantly impacting multiple LinkedIn teams (e.g., topicality, setwise ranking, comments ranking).

LinkedIn, Sunnyvale, CA

Senior Software Engineer - Machine Learning

  • Led the design and implementation of a Gold Content Explore-Exploit (E/E) system based on Thompson sampling and Neural E/E, resulting in a +5.6% increase in Daily Unique Gold Professional Interactions.
  • Identified bottlenecks and designed optimizations, increasing productivity of the E/E stack by ~40%.
  • Worked on fine-tuning machine learning models to improve content recommendation strategies and optimize user engagement.

LinkedIn, Bangalore, India

Senior Software Engineer - AI

  • As a founding member of the Misinformation AI Team, shaped LinkedIn’s misinformation detection strategy and built topic-agnostic misinfo detection models leveraging content, user activity, and network signals.
  • Upgraded and launched a real-time image near-duplicate detection system, significantly improving content moderation processes.
  • Developed a propagation-based viral spam detection library using Graph Neural Networks, strengthening early-detection capabilities for emerging trends and topics.
  • Contributed to models responsible for preventing a unique member impact of ~0.5 million per week due to misinformation.

LinkedIn, Bangalore, India

Applied Research Engineer

  • Spearheaded the launch of new ranking model trained on less noisy data, reducing training data usage by 99.6%, leading to training time reduction by 87.5% while improving CTR@1 by 2.9% and CTR@5 by 1.38%
  • Championed and deployed a novel offline replay model-evaluation pipeline based on NDCG

IBM Research Labs, Bangalore, India

Staff Research Engineer

  • Led the development of an end-to-end Neural Program Induction-based Knowledge Base Question Answering (KBQA) agent, CIPITR, trained with weak supervision using Deep Reinforcement Learning (DRL).
  • Reduced exponential action search space complexity to <0.1% per time-step using beam pruning, reward bonuses, and intermediate rewards, improving model efficiency and scalability.
  • Achieved a F1-score of 58.92% on the CSQA dataset, which was +450% better than the state-of-the-art Neural Symbolic Machines (NSM) F1-score of 10.63%.
  • Filed a patent for a Decentralized Online Multi-Agent Visual Question Answering system, enhancing the robustness of visual QA systems in decentralized environments.

IBM Research Labs, Bangalore, India

Software Engineering Researcher

  • Extended the CIPITR KBQA system to handle noisy datasets without gold query annotations, resulting in the SSRP algorithm, which handled catastrophic forgetting and improved model performance by +24.3% over baselines.
  • Developed a noise-resilient deep-RL algorithm (SSRP) that outperformed state-of-the-art models on noisy datasets, with an F1-score of 72.61%, +5.23% better than NSM.
  • Conducted research on program induction for query answering with minimal supervision, resulting in notable publications and recognition in top conferences like IJCAI and ACL.

IBM Research Labs, Bangalore, India

Intern

  • Developed a touch-based authentication system for mobile devices, using geometric stroke features and vector quantization to optimize computational efficiency and enhance battery life.
  • Achieved an Equal Error Rate (EER) of 17% for KNN-based classification in a real-world scenario.

Education

Indian Institute of Technology Madras, Chennai

Jul '12 - Jul '16

Bachelor of Technology in Electrical Engineering

Indian Institute of Technology Madras, Chennai

Jul '16 - Jul '17

Master of Technology in Electrical Engineering

Projects

IBM Research Labs, Bangalore    (Aug ’17 - Current)

Complex Program Induction for Querying Knowledge Bases in the Absence of Gold Programs  (TACL 2019)(ACL 2019)(Dec ’17 - Jan ’19)

  • Program Induction for Q&A is a paradigm of answering a question by instantiating Operator x Variable sequences
  • The program is executed over a KB and the system receives only the reward obtained as feedback
  • Our operators can encompass all kinds of questions in the CSQA dataset
  • On the CSQA dataset our CIPITR has a F1-score(%) of 58.92 which is +450% improvement than that of the state-of-the-art NSM (10.63)

Neural Program Induction for KBQA Without Gold Programs or Query Annotations  (IJCAI 2019)(May ’18 - Feb ’19)

  • KBQA on a more realistic setting wherein there is additional noise induced by absence of Query annotations and only final answer as supervision
  • The lack of gold Query annotations induces debilitating variance leading to catastrophic forgetting, especially when using RL methods
  • Proposed and realized a noise-resilient deep-RL Algorithm, SSRP that evades noise-induced instability through continual retrospection
  • On the WebQuestionsSP dataset, our SSRP has an overall F1-score(%) of 72.61 and it outperforms state-of- the-art NSM by +5.23%
  • On the noisy CQA-12K dataset we compared F1-Score(%) obtained by our SSRP extensively against SRP (CIPITR), A2C & NSM. On the entire test set, SSRP outperformed the closest baseline SRP by +24.3%

Realistic Online Symbolic Visual Question Answering   (Jan ’19 - Current)

  • The general notion of Symbolic Reasoning VQA methods is to answer a question by inducing neural programs over Operators×Concepts trajectories
  • Currently, SOTA works on VQA focus on offline learning from static datasets where offline learning involves only a fixed set of concepts
  • In a real life settings, unseen concepts can come up both in the referenced datasource(from where concepts are extracted) and questions from users
  • Building a lifelong learning model for VQA inspired from state machine literature
  • Proposed the idea/backbone for enabling continual learning in the given setting and implemented the entire symbolic reasoning pipeline in pytorch framework

IBM Research Labs, Bangalore    (Summer ’15)

Continuous authentication on mobile phones using touchscreen input  (May ’15 - Jun ’15)

  • Goal was to authenticate users based upon touch data
  • Computed 30 geometric feature representations of each stroke
  • Applied vector quantization methods on the stroke data to reduce the computational complexity and optimize data storage, eventually optimizing battery life
  • Achieved 17% Equal Error Rate on just single swipe classification on applying KNN based classifiers

IIT Madras, Chennai    (Jul ’12 - Jul '17)

Language Expansion in Text-Based Games  (Jun ’15 - Sep ’16)

  • Explored interaction based representation learning in the context of learning to play text-based games
  • Designed a Deep-RL based mechanism for expanding an agent’s vocabulary using the vocabulary of agents trained for multiple text-based games
  • Empirically established the utility of embeddings learned by our method by qualitatively analyzing the tSNE embeddings and the transfer learning performance

RL agent to play Flappy Bird game  (Dec ’13 - Jan ’14)

  • Implemented SARSA, SARSA(λ), Q-Learning and TD(λ) reinforcement learning algorithms
  • Obtained a best score of ≈24 which is much better than an average human’s score

Intelligent Ground Vehicle Competition  (May ’16)

  • Member of the image processing team, whose objective was to detect lanes on the ground and avoid obstacles
  • Engineered an algorithm for lane prediction given partial lane detection information and camera induced noise, for assisting in maintaining course

Publications

Complex Program Induction for Querying Knowledge Bases in the Absence of Gold Programs

Saha et al. 2019

TACL 2019, ACL 2019 Oral

Authors:


Neural Program Induction for KBQA Without Gold Programs or Query Annotations

Ansari et al. 2019

IJCAI 2019

Authors:


Language Expansion In Text-Based Games

Ansari et al. 2016

NIPS-DRL Workshop ’16

Authors:

Patents

Hobbies

Skills

Get in Touch