[{"content":" Context # Iris is the conversational assistant in JSI\u0026rsquo;s 4Sight digital intelligence platform. Investigators ask questions in plain language and Iris searches across audio, text, and image evidence, summarizes what it finds, and supports follow-up questions. Everything runs on-premises and air-gapped: no data leaves the customer\u0026rsquo;s site.\nWhat I built # Agentic retrieval and summarization. Led the migration from a proprietary orchestration framework to Agno/LangChain, producing a modular agent system integrated with MCP. Real-time multimodal RAG. Built the pipeline that embeds and retrieves billions of audio, text, and image records, improving search latency and contextual relevance. Evaluation framework. Spearheaded the company-wide evaluation framework, using DeepEval and A/B testing to track relevance, latency, and interaction quality release over release. Scale # Deployed on-premises across customer sites with 100+ GPUs Billions of multimodal records indexed for retrieval Stack # Python PyTorch Agno LangChain MCP DeepEval Kubernetes ","date":"January 2024","externalUrl":null,"permalink":"/projects/iris-ai-assistant/","section":"Projects","summary":"Conversational, on-premises LLM assistant for digital investigations: agentic retrieval and summarization over billions of multimodal records.","title":"Iris: Investigative AI Assistant","type":"projects"},{"content":" Context # Sentry is the always-on detection component of JSI\u0026rsquo;s 4Sight platform. Rather than waiting for an investigator to search, it continuously scans incoming and historical data for high-risk signals and raises alerts with supporting evidence.\nWhat I built # Real-time detection. Designed and deployed a threat detection system that scans high-volume multimodal streams (audio, text, images, video) and flags high-risk activity using anomaly detection and proprietary ML models. Event-driven alerting. Proposed and implemented a Kafka-based, event-driven architecture for time-sensitive alerting workflows, reducing end-to-end notification latency by 33%. Stack # Python Kafka Anomaly detection HPC ","date":"May 2019","externalUrl":null,"permalink":"/projects/sentry-detection-system/","section":"Projects","summary":"Always-on detection over high-volume multimodal data streams, with an event-driven alerting architecture that cut notification latency by 33%.","title":"Sentry: Real-Time Threat Detection","type":"projects"},{"content":" Paper Code Problem # ECG machines are everywhere, inside and outside the clinic, so passive arrhythmia monitoring is possible. But labelled ECG data is scarce, imbalanced, and noisy, and supervised models overfit to those problems.\nApproach # Masked Patch Modelling (MPM): a self-supervised pre-training objective that masks patches of the ECG signal and learns to reconstruct them. Large-scale pre-training on 8.2 million unlabelled ECGs to produce PatchECG, a foundational 1D transformer that can be fine-tuned for any downstream ECG task. Ablations on other remedies for class imbalance, label noise, and over-parameterization. Results # State-of-the-art on standard benchmarks, including PTB-XL multi-label classification, and new benchmarks on the largest, highest-quality multi-label dataset to date. Outperforms the previous state-of-the-art using 1/5 of the compute while increasing model capacity 14×. Significantly outperforms a state-of-the-art 2D vision transformer. More than 2% improvement under class imbalance, label noise, and over-parameterization. ","date":"September 2024","externalUrl":null,"permalink":"/projects/automated_ecg/","section":"Projects","summary":"Master’s thesis, published in Scientific Reports (2026). A 1D transformer pre-trained on 8.2M unlabelled ECGs with Masked Patch Modelling; state-of-the-art on PTB-XL at 1/5 the compute.","title":"PatchECG: Self-Supervised ECG Arrhythmia Detection","type":"projects"},{"content":" Code An implementation of the min-max and min-max+ algorithms for distributed leader election in a ring, built for the graduate course CSI 5308. The simulation supports any ring size and any number of originators.\n","date":"May 2023","externalUrl":null,"permalink":"/projects/leader_election/","section":"Projects","summary":"Implementation and visualization of the min-max and min-max+ leader election algorithms for any ring size and number of originators.","title":"Distributed Leader Election in a Ring","type":"projects"},{"content":"","date":"March 2026","externalUrl":null,"permalink":"/tags/medical-ai/","section":"Tags","summary":"","title":"Medical AI","type":"tags"},{"content":"I\u0026rsquo;m a Machine Learning Engineer who builds agentic AI systems and the frameworks that evaluate them at scale.\nAt Layer 6 AI (TD Bank), I\u0026rsquo;m on the Agentic team, where we build long-running agents that take on complex tasks within a highly regulated environment. I built the automated evaluation framework used to assess agentic conversations across the bank, cutting the time to deploy new evaluations from roughly six months to a few weeks.\nPreviously at JSI, I built an agentic harness that lets LLMs call internal APIs through MCP and search billions of multimodal records (audio, text, and image), deployed on-premises across customer sites with 100+ GPUs.\nMy Master\u0026rsquo;s research at Carleton University, published in Scientific Reports, introduced PatchECG: a self-supervised 1D transformer pre-trained on 8.2 million unlabelled ECGs that achieved state-of-the-art results in arrhythmia detection.\n","date":"March 2026","externalUrl":null,"permalink":"/","section":"Mitchell Chatterjee","summary":"","title":"Mitchell Chatterjee","type":"page"},{"content":"","date":"March 2026","externalUrl":null,"permalink":"/publications/","section":"Publications","summary":"","title":"Publications","type":"publications"},{"content":"","date":"March 2026","externalUrl":null,"permalink":"/tags/self-supervised-learning/","section":"Tags","summary":"","title":"Self-Supervised Learning","type":"tags"},{"content":"","date":"March 2026","externalUrl":null,"permalink":"/tags/","section":"Tags","summary":"","title":"Tags","type":"tags"},{"content":"Mitchell Chatterjee, Adrian D. C. Chan, Majid Komeili · Scientific Reports 16, 11793 (2026) · DOI 10.1038/s41598-026-41549-5\nJournal Open access (PMC) Code Project summary Abstract # Cardiovascular diseases are the leading cause of death worldwide. With electrocardiogram (ECG) machines becoming more accessible, passive monitoring for arrhythmia detection is now possible. This work highlights the importance of self-supervised learning in detecting arrhythmias by leveraging large-scale unlabelled ECG data to improve performance and reduce overfitting to class imbalance and noise. We propose Masked Patch Modelling (MPM) and use 8.2 million unlabelled ECGs for self-supervised pre-training, introducing PatchECG, a 1D Transformer model that can be fine-tuned for various ECG tasks. PatchECG achieves state-of-the-art results on standard datasets, including PTB-XL multi-label classification, and sets new benchmarks on the largest and highest-quality multi-label dataset to date. Compared to existing methods, PatchECG is five times more computationally efficient while increasing model capacity by a factor of 14. We also compare the 1D PatchECG model to a state-of-the-art 2D vision Transformer, HeartBEiT, and observe significantly higher performance. Finally, ablation studies reveal a 2% performance improvement in handling class imbalance, label noise, and over-parameterization. These findings demonstrate the potential of self-supervised learning in advancing automated arrhythmia detection.\nOther versions # Preprint: TechRxiv, 2024 · DOI 10.36227/techrxiv.172866031.13011158/v1 Master\u0026rsquo;s thesis: Carleton University, 2024 · DOI 10.22215/etd/2024-16267 Cite # @article{chatterjee2026patchecg, title = {Toward robust automated cardiovascular arrhythmia detection using self-supervised learning and 1-dimensional vision transformers}, author = {Chatterjee, Mitchell and Chan, Adrian D. C. and Komeili, Majid}, journal = {Scientific Reports}, volume = {16}, pages = {11793}, year = {2026}, doi = {10.1038/s41598-026-41549-5} } Cover image: \"ECG and CVP Curves\" by Stefan Bellini, public domain (CC0) via Wikimedia Commons.\n","date":"March 2026","externalUrl":null,"permalink":"/publications/patchecg/","section":"Publications","summary":"Mitchell Chatterjee, Adrian D. C. Chan, Majid Komeili · Scientific Reports, 2026","title":"Toward robust automated cardiovascular arrhythmia detection using self-supervised learning and 1-dimensional vision transformers","type":"publications"},{"content":"Anne Kim, Mitchell Chatterjee, Alla Iansavitchene, Majid Komeili, Adrian D.C. Chan, Homer Yang, Jason Chui · British Journal of Anaesthesia 135(3): 561–570, 2025 · DOI 10.1016/j.bja.2025.05.037\nJournal Open access (PMC) Abstract # Background. Perioperative electrocardiographic monitoring can offer immediate detection of myocardial ischaemia, yet its application in perioperative and remote monitoring settings is hampered by frequent false alarms and signal contamination. We performed a scoping review for the current state of artificial intelligence (AI) in perioperative ECG interpretation.\nMethods. A literature search in Ovid MEDLINE, EMBASE, Compendex, and CINAHL databases was performed from inception to May 10, 2023. All original research of ECG monitoring for myocardial ischaemia, myocardial infarction, or both was included.\nResults. A total of 182 original research articles published between 1991 and 2023 were included. Most studies (n=132) used a pre-existing ECG database to develop AI algorithms retrospectively, and the rest did not specify their sources. Processing filters were used in 58% of the studies to remove ECG noises/artifacts before AI algorithm development. Amongst the AI technologies used, ResNet demonstrated the highest median sensitivity, precision, and specificity at 98.4%, 99.8%, and 99.1%, respectively. There are only five studies with intermittent prospective ECG collection on ST-segment elevation myocardial infarction. No studies prospectively collected continuous ECG perioperatively, with frequent false alarms and signal contamination.\nConclusions. AI technology can achieve high diagnostic accuracy for myocardial ischaemia detection in clean intermittent electrocardiograms. However, almost all these algorithms were developed from a few open-source clean ECG databases without testing on \u0026rsquo;noisy data\u0026rsquo;, which limited their clinical applicability in the perioperative setting where signal contamination is frequent. AI algorithms on perioperative electrocardiography, tested in a noisy perioperative and remote monitoring environment, including wearable devices, are needed.\nCite # @article{kim2025perioperative, title = {Artificial intelligence for electrocardiographic diagnosis of perioperative myocardial ischaemia: a scoping review}, author = {Kim, Anne and Chatterjee, Mitchell and Iansavitchene, Alla and Komeili, Majid and Chan, Adrian D. C. and Yang, Homer and Chui, Jason}, journal = {British Journal of Anaesthesia}, volume = {135}, number = {3}, pages = {561--570}, year = {2025}, doi = {10.1016/j.bja.2025.05.037} } Cover image: hospital operating room, public domain (CC0) via rawpixel.\n","date":"September 2025","externalUrl":null,"permalink":"/publications/perioperative-ecg-ai/","section":"Publications","summary":"Anne Kim, Mitchell Chatterjee, Alla Iansavitchene, Majid Komeili, Adrian D.C. Chan, Homer Yang, Jason Chui · British Journal of Anaesthesia, 2025","title":"Artificial intelligence for electrocardiographic diagnosis of perioperative myocardial ischaemia: a scoping review","type":"publications"},{"content":"","date":"September 2025","externalUrl":null,"permalink":"/tags/ecg/","section":"Tags","summary":"","title":"ECG","type":"tags"},{"content":"","date":"September 2025","externalUrl":null,"permalink":"/tags/scoping-review/","section":"Tags","summary":"","title":"Scoping Review","type":"tags"},{"content":"Production systems I\u0026rsquo;ve built at JSI, and research from my Master\u0026rsquo;s.\n","date":"September 2024","externalUrl":null,"permalink":"/projects/","section":"Projects","summary":"","title":"Projects","type":"projects"},{"content":"","date":"September 2024","externalUrl":null,"permalink":"/tags/time-series/","section":"Tags","summary":"","title":"Time-Series","type":"tags"},{"content":"","date":"September 2024","externalUrl":null,"permalink":"/tags/transformers/","section":"Tags","summary":"","title":"Transformers","type":"tags"},{"content":"","date":"January 2024","externalUrl":null,"permalink":"/tags/agents/","section":"Tags","summary":"","title":"Agents","type":"tags"},{"content":"","date":"January 2024","externalUrl":null,"permalink":"/tags/evaluation/","section":"Tags","summary":"","title":"Evaluation","type":"tags"},{"content":"","date":"January 2024","externalUrl":null,"permalink":"/tags/llms/","section":"Tags","summary":"","title":"LLMs","type":"tags"},{"content":"","date":"January 2024","externalUrl":null,"permalink":"/tags/rag/","section":"Tags","summary":"","title":"RAG","type":"tags"},{"content":"","date":"May 2023","externalUrl":null,"permalink":"/tags/distributed-algorithms/","section":"Tags","summary":"","title":"Distributed Algorithms","type":"tags"},{"content":"","date":"May 2019","externalUrl":null,"permalink":"/tags/anomaly-detection/","section":"Tags","summary":"","title":"Anomaly Detection","type":"tags"},{"content":"","date":"May 2019","externalUrl":null,"permalink":"/tags/kafka/","section":"Tags","summary":"","title":"Kafka","type":"tags"},{"content":"","date":"May 2019","externalUrl":null,"permalink":"/tags/streaming/","section":"Tags","summary":"","title":"Streaming","type":"tags"},{"content":" Work # Machine Learning Engineer Mar 2026 – Present Layer 6 AI · TD Bank Built the automated evaluation framework on the Agentic AI team, now used to evaluate agentic conversations across the entire bank. Cut deployment time for new automated evaluation solutions from roughly six months to a few weeks by standardizing the evaluation pipeline and making evaluators configurable. AI / Machine Learning Engineer Dec 2024 – Feb 2026 JSI Led the migration from a proprietary orchestration framework to Agno/LangChain, enabling a modular, LLM-based agentic retrieval and summarization system integrated with MCP; deployed on-premises across customer sites with 100+ GPUs. Built a real-time RAG pipeline to embed and retrieve billions of multimodal records (audio, text, image), significantly improving search latency and contextual relevance. Spearheaded the company-wide evaluation framework, using DeepEval and A/B testing to measure relevance, latency, and interaction quality. Junior AI / Machine Learning Engineer May 2019 – Sep 2022 JSI Designed and deployed a real-time threat detection system that scanned high-volume multimodal data streams (audio, text, images, video) to detect and alert on high-risk activity using anomaly detection and proprietary ML models. Proposed and implemented a Kafka-based event-driven architecture for time-sensitive alerting, reducing end-to-end notification latency by 33%. Education # Master of Computer Science (Thesis), Machine Learning 2022 – 2024 Carleton University GPA 4.0 / 4.0. Thesis: Toward Robust Automated Cardiovascular Arrhythmia Detection using Self-supervised Learning and 1D Vision Transformers.\nCourses: Advanced Machine Learning, Natural Language Processing, Deep/Reinforcement Learning.\nHonours BSc Computer Science, AI Specialization 2016 – 2020 University of Ottawa GPA 4.0 / 4.0, summa cum laude. Skills # Machine learning\nTransformers \u0026amp; LLMs RAG \u0026amp; vector search Agents \u0026amp; MCP Self-supervised learning Parameter-efficient fine-tuning Time-series LLM evaluation Engineering\nPython PyTorch LangChain / Agno DeepEval Docker \u0026amp; Kubernetes Kafka HPC / GPU SQL C# JavaScript Languages: English (fluent), French (intermediate)\n","externalUrl":null,"permalink":"/experience/","section":"Mitchell Chatterjee","summary":"","title":"Experience","type":"page"}]