Hey, I'm

Prahar Ijner

ML engineer and researcher

About

Over the past 5+ years I've been building production ML systems in domains where the quality bar is high — medical imaging, autonomous systems, Earth observation. I did my master's in computer engineering at the University of Waterloo, and the thing that stuck most was the research side: digging into failure modes, finding where systems quietly break, and catching it before they ship. That's the thread through most of the work below.

Selected research

Reliability and evaluation of ML systems — finding the slices, inputs, and conditions where models fail.

NeurIPS 2025 Active learning LLM evaluation

Active Slice Discovery in Large Language Models

*Minhui Zhang, *Prahar Ijner, Yoav Wald, Elliot Creager · *equal contribution

LLMs fail systematically on hidden subsets of data — a demographic, a phrasing, a topic. Finding those "error slices" usually takes heavy manual labeling. We show you can find them cheaply.

We formalize Active Slice Discovery: actively grouping errors likely to share a failure pattern, then using limited annotator access to confirm the slice. On toxicity classification, uncertainty-based active learning recovers human-defined slices while significantly outperforming baselines.

2–10%

of slice labels needed to recover a slice

uncertainty sampling wins

arXiv:2511.20713

RAG Hallucination detection Robustness

Hallucination Detection Under Retrieval Corruption

Prahar Ijner · Graduate research, University of Waterloo

RAG hallucination detectors are trained on clean retrieval. Real retrieval is noisy. Do the detectors survive? I stress-tested them across 80 degradation conditions to find out.

I benchmarked a DeBERTa NLI cross-encoder against a purpose-built detector on FEVER, sweeping four corruption strategies across five severities and two retrieval modes. The finding: detectors are largely robust to reordering and contradictory evidence, but collapse under distractor injection — meaning retrieval recall, not detector robustness, is the real bottleneck.

corruption conditions evaluated

0.043

NLI calibration error (ECE) vs 0.113 baseline

recall is the real gap

Code & report

Backdoor attacks Adversarial ML Tabular data Network security

Backdoor Learning & Defenses for Safety-Critical ML

Prahar Ijner · Graduate research, University of Waterloo

Backdoor triggers embedded in just 2% of training data produce >93% attack success while leaving F1 scores nearly unchanged. Standard validation is blind to the compromise.

I trained DNN, TabNet, XGBoost, and Random Forest on the ToN-IoT network intrusion dataset, injecting composite backdoor triggers at 2%, 5%, and 10% poisoning rates. I benchmarked three existing defenses — SHAP attribution deviation, STRIP, and Activation Clustering — and proposed DeltaSHAP, which detects poisoning by comparing feature attributions between a clean reference model and the suspect model. DeltaSHAP was the only method that held up consistently across both neural and tree-based architectures, where other defenses break down.

>93%

backdoor success rate at just 2% data poisoning

model architectures benchmarked (DNN, TabNet, XGBoost, RF)

know your threat model

GitHub

More projects

Things I've built to learn something, win something, or scratch an itch.

CUCAI 2025 award

Satellite methane detection

Led a team building real-time methane leak detection on hyperspectral satellite imagery. Won the AI for Environmental Sustainability award at CUCAI 2025.

WatAI · SatML →

Quantum autoencoders

Comparing quantum and classical autoencoders for image-patch compression, built with Qiskit and TensorFlow.

GitHub →

Feed-forward learning

A from-scratch classification algorithm that avoids backpropagation entirely, stacking layers of kernel-similarity "alpha" scores to build a decision boundary, benchmarked against standard classifiers across 9 datasets.

GitHub →

Genetic ensemble orchestration

Undergraduate thesis: using genetic algorithms to orchestrate deep-learning ensembles for image classification, benchmarked across datasets.

GitHub →

Experience

Senior Data Scientist

Jan 2025 — Present

Radformation · radiation oncology software

Data Scientist

May 2024 — Jan 2025

Radformation · radiation oncology software

Machine Learning Engineer

Jan 2022 — May 2024

Limbus AI · automated contouring for radiotherapy — acquired by Radformation (2024)

Software Developer

May 2021 — Jan 2022

Kraken Robotics · autonomous underwater vehicles

Want the role-by-role detail? It's in my résumé — just ask.

Languages

PythonC++RustCUDASQL

Modeling

PyTorchTransformersCLIPDiffusionVLMsONNX

LLMs & agents

LangGraphLangChainRAGvLLMFAISSPinecone

MLOps & infra

AWSGCPDockerKubernetesMLflowW&BRayGrafanaSLURM / DDP

Focus

LLMsRobustnessActive learningEvaluationOOD / drift

Get in touch

If something here caught your interest — a paper, a project, or a problem you're chewing on — I'd be glad to hear from you.

Name

Message

LinkedIn GitHub Scholar

replies faster than you'd expect