Incoming M.Sc. student in Computing Science with GTA, GRA, and GRAF appointments supporting graduate teaching and research in software engineering and dependable AI.
Begin graduate research with the U-A-Goose software engineering group under Dr. Zhou Yang
Support University of Alberta teaching and research through GTA and GRA appointments with GRAF funding
Graduate Research · Teaching · Software Engineering · Dependable AI
Remote summer research internship at the University of Illinois Urbana-Champaign through UIUC++ Summer Research in Software Engineering (SRSE) 2026 with Professor Tianyin Xu's group, focused on software and systems reliability research.
Working on software engineering research with an emphasis on reliable systems and rigorous empirical evaluation
Reading systems and software reliability literature, prototyping research ideas, and preparing artifacts for reproducible evaluation
Software Engineering Research · Systems Reliability · Empirical Evaluation · Reproducibility
Backend engineer on IQVIA's KPI Library, the dynamic reporting layer within Orchestrated Analytics, working across configuration, dashboard, analysis, and export workflows for healthcare analytics. Built and maintained C#/.NET services, data-access components, automated test infrastructure, and a schema-grounded AI copilot with supporting evaluation, retrieval, security, and observability infrastructure.
Built and maintained C#/.NET services and shared components across four KPI Library workflow areas—configuration, dashboard, export, and gateway; used EF Core with SQL Server for configuration data, MongoDB for dashboard caching, and AWS S3 in export workflows
Refactored repeated filter-resolution logic into reusable query components, consolidating 6 duplicated implementations into 2 shared components; centralized validation and business rules and added regression scenarios for combined filters and edge cases
Added in-memory caching to targeted .NET read paths and improved interaction with the MongoDB-backed dashboard cache, reducing MongoDB reads per warm request from 4 to 1 and p95 latency from 420 ms to 340 ms across 500 equivalent warm requests per benchmark run
Additional responsibilities and evidence
Extended two export paths—CSV and presentation exports—with regional-formatting and commentary support; increased automated high-risk export scenarios from 28 to 46 and reduced pipelines requiring a flaky rerun from 12 of 100 to 4 of 100 comparable export-test runs by controlling fixtures, locale, and browser state and normalizing volatile file metadata before hashing stable content
Implemented server-side axis-break resolution for KPI charts by computing adjacent gaps in ordered KPI values and translating qualifying intervals into Highcharts yAxis.breaks, keeping visualization behavior consistent across clients
Reorganized three CI stages, build, test, and packaging; ran independent test suites in parallel and reused build artifacts between stages, reducing median successful-pipeline duration from 24.5 to 20.3 minutes across 25 comparable runs before and 25 after the change
Contributed OpenTelemetry/Jaeger instrumentation and correlation identifiers across 8 named API, database, and cache operations in three workflow areas—dashboard, configuration, and export—making cross-service calls and database latency visible during debugging
Designed a schema-grounded dashboard copilot that compiled analyst requests into versioned, permission-scoped operations for KPI selection, filtering, comparison, visualization, commentary, and export; validated and previewed each plan before approval, executed it idempotently with concurrency and rollback safeguards, then verified the final state and recorded an audit trail, improving exact requested-state completion from 65% to 87% and reducing invalid committed changes from 8.1% to 0.9% on a 350-workflow held-out subset of the agent benchmark
Built a 400-workflow agent evaluation platform with trace replay, deterministic state and business-rule graders, numerical and export assertions, human-calibrated narrative scoring, multi-trial execution, and CI release gates; caught more than 20 prompt, model, retrieval, and tool-schema regressions before rollout and reduced candidate-model qualification from five days to one
Developed an evidence-grounded KPI investigation assistant combining deterministic calculations, anomaly detection, hierarchical driver analysis, and source-linked narratives; increased numerical claims matching computed evidence from 91% to 98.7%, reduced unsupported analytical statements from 12% to 3%, and cut median investigation time from 13 to 7 minutes on an expert-reviewed task set
Built a policy-enforcing execution gateway with capability-scoped authorization, tenant and market propagation, sensitive-data redaction, approval workflows, untrusted-content isolation, and immutable provenance; observed zero unauthorized executions across 5,000 adversarial cases while keeping benign-task false blocks below 2%
Developed tenant-, permission-, and version-aware hybrid retrieval over KPI definitions, calculation rules, dashboard templates, and operational knowledge using lexical and embedding retrieval, reranking, effective-date filters, structured lookups, versioned citations, and calibrated abstention; improved Recall@5 from 0.76 to 0.93, citation precision from 84% to 97%, and reduced stale KPI explanations from 12% to 3% on a held-out retrieval evaluation
Built an observable, provider-neutral model gateway with versioned prompts and models, complexity-aware routing, scoped caching, cost budgets, bounded retries, circuit breakers, fallbacks, shadow evaluation, and end-to-end AI traces; reduced p95 validated-response latency from 17 to 9 seconds and cost per successful workflow by 42%, while keeping candidate-model task success within one percentage point of the baseline model
Received the IQVIA Impact Program - Silver award in May 2025
