Current focus
Applied AI, NLP, and LLM systems
Mohaiminul
Md Mohaiminul Islam
Senior Data Scientist · Workers' Compensation Board (WCB), Edmonton
Applied AI that stands up to review.
70M+ documents · ~5K/day scored · ~91% F1 across 14 classes
I build reviewable AI for regulated document work: PII detection, Databricks RAG, and governed releases.
Proof at a glance
Fit for regulated AI delivery.
Current work, measured results, and platform depth before the case studies.
Strongest signal
Reviewable PII detection across WCB's historical claims corpus
Working pattern
Model work paired with Databricks pipelines, evaluation, and human review
Helpdesk virtual agent
Approved for company-wide rollout to ~3,500 employees
Employee-facing Databricks RAG app; a two-week helpdesk-team trial found ~70% of answers helpful, with remaining gaps traced to connecting more internal sources.
Forms workflow POC
Evaluation shows form returns can drop from ~40% to ~10%
Medical report-generation agents with human review; prototype evaluated on historical data, now under departmental review.
Agentic maintenance
Monitors daily production jobs and stages fixes for human approval
Built on Databricks LLM endpoints (Claude), the agentic harness diagnoses failures and validates fixes on sandboxed data samples; later extended to automated code review.
Claim-duration model
3-class XGBoost model in production, scoring ~1K claims/day
Chosen over NLP/Transformer alternatives after benchmarking accuracy, speed, and cost; monitored weekly with retraining alerts.
Annotation automation
Avoided an estimated $25K in manual labeling cost
AWS Lambda workflow for LLM data generation, validation, and annotation at Blue Guardian Canada.
Data preparation
About 2M records prepared; model-prep time down ~30%
Config-driven PySpark preprocessing and embedding pipelines at Servier Canada.
Featured work
Start with the WCB PII system.
The flagship work shows the pattern: reviewable AI, measurable output, and delivery inside a regulated document workflow. The next two notes add scale and research depth.
Flagship case · Workers' Compensation Board (WCB) · 2023 - Current
Reviewable PII detection for claim documents
Designed and operationalized end-to-end PII detection for a roughly 70M-document historical claims corpus and ~5K new documents/day, combining prompt-routed LLM inference, deterministic validation, post-processing, and reviewable evidence.
Problem
Claim records arrived as inconsistent PDF-extracted text, but downstream work needed reliable detection across 14 PII/entity classes including names, identifiers, dates, addresses, and other sensitive fields.
Evaluation
Measured precision, recall, and F1 against labeled data, then reviewed outputs for evidence quality, overlap resolution, and traceability before downstream use.
Decisions
- Routed documents by type and text shape so extraction stayed specific to the document pattern.
- Combined LLM extraction with regex, dictionary, structural-data matching, and span repair to clean up entity evidence.
- Kept one YAML-driven workflow contract so local and Databricks runs produced comparable artifacts, MLflow tracking, and release-gate evidence.
Review boundary
Weak spans, unsupported identifiers, runtime drift, or outputs without traceable evidence and run logging were treated as failure states.
Evidence stackSenior Data Scientist2023 - Current
ToolsPythonLLMsAzure DatabricksMLflow
Confidential-work summary. Internal document examples, prompt content, and sensitive process details stay out of scope.
02
Blue Guardian Canada Inc. · 2023
A 29-class mental-health text classifier at ~92% F1
Managed 4 direct-report data science interns delivering a 29-class mental-health text classifier, pairing transformer and LLM modeling with LLM-generated synthetic text and automated annotation.
03
Publication-grounded research · 2015 - 2023
Privacy-preserving biomedical prediction
Built privacy-preserving deep learning for drug sensitivity prediction on ~11,000 scRNA-seq samples and contributed publication-backed biomedical machine learning research.
Review loop
From messy documents to reviewed output.
The flagship operating pattern: narrow AI assistance, human review, data iteration, and measured outcomes.
Messy input
Text-heavy forms and internal workflows
PII, support requests, and medical report-generation agents.
Narrow AI assist
Extraction, redaction, and form preparation
LLM work stays scoped to steps staff can check.
Staff review
Human checks remain in the operating path
Outputs are useful only when staff can inspect them.
Evaluated workflow
Form returns can drop from ~40% to ~10%
Historical-data evaluation of the medical report-generation agents, now under departmental review.
Databricks release gatesPII detectionEmployee-facing RAGForms workflow POC
Deep dives
Proof pages keep the details inspectable.
Each page expands one approved system story with the outcome, decision record, verification path, and boundaries a reviewer can trust.
Experience
Experience and education.
Mohaiminul is a Senior Data Scientist in Edmonton, building reviewable LLM and NLP systems for claim-document workflows at WCB. Earlier roles add transformer team leadership, annotation automation, and data-platform work, backed by graduate research in privacy-aware biomedical machine learning.
2023 - Current
Workers' Compensation Board (WCB)
Senior Data Scientist
PII detection at document-corpus scale, Databricks pipelines, employee-facing RAG, agent-assisted forms, and an agentic maintenance harness for regulated claim workflows. Managed 2 interns, mentored 5 junior data scientists, supported 5 hiring processes, and presented recommendations to the CTO and director-level stakeholders.
2023
Quantolio
Senior Data Scientist
Python refactoring, Vision Transformer analysis, and a Streamlit portfolio-analytics prototype with time-series analysis and reinforcement-learning experiments for portfolio optimization.
2023
Blue Guardian Canada Inc.
Lead Data Scientist, ML/NLP
29-class mental-health text classification, synthetic-data strategy, annotation automation, and intern-team leadership.
2021 - 2022
Servier Canada
Data Scientist Intern, Mitacs
Prepared about 2 million unstructured records, cut model-preparation time by ~30%, and supported generative chemistry model work.
2015 - 2023
University of Manitoba
Research Assistant
Privacy-preserving biomedical ML, published research, and evaluation-led modeling across sensitive datasets.
Earlier work includes a data-science internship at Sightline Innovation and teaching in computer science.
Education
- PhD Candidate in Computer Science (candidacy completed), University of Manitoba
- MSc in Computer Science, University of Manitoba
- BSc in Computer Science and Engineering, University of Chittagong
Capability stack
Applied AI, data platforms, and technical leadership.
The practical stack behind the resume and case studies.
Applied AI systems
- Reviewable LLM workflows
- PII and entity detection
- RAG
- AI agents
- LangChain
- LlamaIndex
- Databricks Vector Search
- Hugging Face + vLLM
- Human review loops
Data platforms
- Azure Databricks
- PySpark
- ETL pipelines
- SQL
- Unity Catalog
- Model monitoring
- Pandas
- NumPy
- GCP Vertex AI notebooks
ML engineering
- Python
- PyTorch
- XGBoost
- Transformers (BERT)
- Clustering and unsupervised ML
- Reinforcement learning
- LLM-as-judge evaluation
- TensorFlow
- MLflow
- DVC
- FastAPI
- Docker
Technical leadership
- 4 direct-report interns (Blue Guardian)
- Junior data scientist mentoring
- Cross-functional coordination
- Reviewable system design
Contact
If staff can't check it, it doesn't ship.
Focus: reviewable LLM/NLP systems, Databricks pipelines, PII detection, and regulated document workflows in Canada.
Md Mohaiminul Islam · Edmonton, Alberta, CanadaApplied AI and NLPRegulated workflows