Pedram Agand

Vancouver, BC (Hybrid)

• On Microsoft's Turing team, advance the AI capabilities of Microsoft 365 Copilot to make it more capable, reliable, and useful, in collaboration with leading partners
• Develop and evaluate advanced AI systems, focused on agentic workflows and large language model orchestration
• Work across the full applied-science lifecycle: defining research questions, designing evaluation datasets, analyzing production telemetry, running controlled experiments, and translating results into product and engineering decisions
• Design rigorous evaluations for complex AI capabilities and analyze model quality, grounding, reliability, latency, and resource efficiency
• Turn ambiguous technical findings into reproducible recommendations and actionable product changes with scientists, engineers, and product partners

Vancouver, BC

• On the Alexa team, led category-agnostic implicit NER with negative sampling to decouple extraction from classification - improved accuracy by 7.2%, reduced missing-entity by 68%, eliminated false positives
• Designed self-consistency mechanism and entity-resolution pipeline cutting manual annotation time by 85%
• Built enterprise-scale AWS monitoring system that auto-triggers LLMaJ integration tests on a gating dataset; live dashboards detect drift and flag configuration regressions
• Owned cross-functional deployment of RFC-compliant content-integrity models; guardrails for toxicity, hallucination, and prompt leakage - 98% safety-benchmark pass rate without latency degradation
• Engineered deterministic post-processing and resilient parsers (regex → robust) for temporal intent classification with integrated A/B testing, restoring extraction accuracy and downstream API interoperability

Burnaby, BC · pagand/AITutor_SeqModeling ↗

• Trained BERT for score prediction with contrastive learning - improved accuracy by 21%
• Designed and deployed an agent-based LLM orchestrator (Llama3 + vector DB + LangGraph) for personalized hints; controlled evaluation improved metric_grade 72.7% and engagement 65.1%
• Built LLMaS for synthetic data generation statistically matching real-learner profiles; adaptive treatment improved BKT probability by 19.6% and reduced attempts per question 4.7% vs. control

Vancouver, BC · pagand/ragprop ↗

• Fine-tuned a self-corrective RAG agent (Phi3 + QLoRA + LangGraph) to extract information from meeting notes and transcripts - reduced staff data-collection time by ~6 hours/week

Burnaby, BC · pagand/NRC ↗

• Implemented ETL pipelines to analyse 2 years of ferry operational data with DNN and ensembles
• Led and mentored 2 USRA students and 1 intern in prototyping DL models for industry-scale data - resulted in 2 published research papers; communicated results to non-technical stakeholders

Vancouver, BC

• Research: estimated density ratio via attention-based network and propensity score estimator in finance
• Engineering: CI/CD pipelines, Slurm cluster, foundation models, MLflow, CR/PR, unit tests

Vancouver, BC · pagand/Eco-Light ↗

• Developed EcoLight, a reward-shaping DRL framework for traffic signal control - reduced CO₂ emissions by 11.3% while limiting travel-time increase to 2.7% (NeurIPS-W 2021)

Burnaby, BC · pagand/human_prediction ↗

• Led development of a particle-filter Bayesian predictor for human navigation, replacing brute-force search - reduced multi-step prediction errors by up to 44% in simulation and ~33% in real-world (ICRA 2022)