Research
Papers and active work
4 publications: 2 journal articles and 2 conference papers. The doctoral work covers convergence of gradient descent and SGD for deep linear networks. The applied work covers anomaly detection and fuel forecasting on live infrastructure. Every paper below links to the published version.
Google Scholar profile, with citation counts →Publications
Convergence of gradient descent for learning linear neural networks
Springer Nature: Advances in Continuous and Discrete Models, 2024
Proves that gradient descent converges to a critical point of the square loss for deep linear networks, under an explicit step-size bound that stops shrinking exponentially as the network deepens. For 2-layer networks you get global convergence for almost all initializations; for 3 layers and up, convergence to a global minimum on the fixed-rank manifold. The bound was then checked on tanh networks and held.
Analysis of stochastic gradient descent for learning linear neural networks
SampTA 2023
Analyzes discrete SGD on nonconvex objectives directly, without assuming Lipschitz-continuous gradients, which neural network losses do not have. Bounds the iterate dynamics to control exploding gradients and proves convergence under decreasing and adaptive learning rate schedules.
Anomaly detection in power generation plants using machine learning and neural networks
Applied Artificial Intelligence, 2020
An anomaly detection framework for diesel generator data at telecom base stations. It weighs accuracy against fairness, so an operator can read why a site was flagged before acting on it.
Predicting fuel consumption in power generation plants using ML and neural networks
ICECET 2021
Fuel consumption models for telecom base stations. This is the forecasting work behind the deployed warning system that helped account for 84,617 liters of diesel.
Active work
Four-stage hybrid RAG for scientific QA
SPECTER2 dense retrieval trained on the citation graph, BM25 sparse search, Reciprocal Rank Fusion, ColBERT v2 reranking, and a CRAG gate over 47,810 QASPER chunks. Latest run on 84 questions: ALCE Citation Precision 0.8373, Faithfulness 0.7150. Context Recall is the open bottleneck, and the ColBERT v2 upgrade targets it.
LLM reasoning and agentic AI evaluation
AI Research Consultant at Outlier AI and Mercor since October 2025. I find where a model's reasoning breaks on hard STEM problems and write the correction that becomes the training signal, grade agents on tool choice and long-task behavior, and design the rubrics that keep a distributed group scoring the same way.