Fraud-detection MLOps
Real-time card-fraud scoring on data where 99%+ of rows are innocent. Leakage-guard tests, threshold-sensitivity suite (which caught a real all-negative-batch crash before production did), Prometheus, Terraform → Cloud Run.
A dossier, not a pitch — every number here links to the run that produced it.1
Models in production. Not in notebooks.
I'm MD Minhazur Rahman — a machine-learning engineer & automation specialist. I build forecasting, fraud-scoring and streaming systems that are typed, tested, monitored, and live on the cloud — then I publish the evidence.
3+ years data & infrastructure background · MSc Data Science (Merit), University of Greenwich, London · dissertation published with a DOI4 · based in Chattogram, Bangladesh — open to on-site, remote and relocation roles.9
Exhibit A · Live services
Fraud scoring, live
Card-fraud detection on heavily imbalanced data — leakage-guarded features, threshold tuned on a held-out set, Prometheus metrics, Terraform-provisioned.
fraud-detection-api…run.app →Demand forecasting, live
Hourly per-SKU retail forecasts with the drivers behind them — MLflow-tracked training, Evidently drift monitoring, keyless CI/CD via Workload Identity Federation.
retail-forecasting-api…run.app →Both scale to zero between visitors — a deliberate $0/month deployment. This page pings them the moment it loads, so by the time you read this they are usually already warm; on an unlucky cold start the first request takes ~30 s.3 Patience is cheaper than idle compute.
−40.8%2
mean absolute error versus a seasonal-naïve baseline — the honest yardstick most forecasting portfolios skip. If a model can't beat the dumb baseline, it doesn't earn a deploy.
Exhibit B · Selected work
Selected work
six systems, all publicRetail forecasting pipeline
End-to-end demand forecasting: Docker, MLflow, FastAPI, Evidently drift monitoring, keyless CI/CD. Chronological validation only — no time-travel in the splits.
Streaming fraud scorer
Kafka/Redpanda stream processor: schema validation into a dead-letter queue, alerts topic, Grafana dashboards-as-code. CI deploys the whole stack onto a kind Kubernetes cluster and replays 500 transactions end-to-end.
500/500 scored in k8s CI6
sourcethe debugging story
Orchestrated retraining platform
Dagster assets build a dbt star schema on DuckDB, retrain a challenger weekly on 1M+ real transactions, and promote it only past a champion gate. Model promotion as code, not hope.
A/B experimentation toolkit
CUPED variance reduction and always-valid sequential testing (mSPRT) — with property-based tests that verify the false-positive guarantees instead of assuming them.
Invoice automation
OCR → structured extraction → validation rules → DuckDB ledger, with a human-in-the-loop review queue for everything below the confidence bar. The flagship of a four-repo automation suite.
deterministic core, LLM optional
sourcefull suite
Twenty-plus repositories in total — tutorials and research artifacts included — at github.com/minhazda.
Exhibit C · Practice
Practice
what I actually doProduction ML engineering
Models served as real APIs on real infrastructure. Infrastructure as code, keyless deploys, metrics and drift monitoring from day one — because a model without observability is a rumor.
fastapi · docker · terraform · cloud run · prometheus · grafana · evidently · mlflow
Forecasting & applied ML
Demand forecasting, fraud detection, experimentation. Honest baselines first; models earn their deploy by beating them on chronologically-split, leakage-guarded data.
lightgbm · scikit-learn · time series · CUPED / mSPRT · pandas
Automation & agentic tooling
Human-in-the-loop automations that do real work — document intake, ticket triage, RPA reconciliation, workflow orchestration. Deterministic cores; LLMs optional and evaluated, never load-bearing by default.
ocr · playwright · n8n · langgraph · pydantic · duckdb
Product delivery, end to end
Designed, built and run Untemplated Studio solo — a productised web-build service with a 48-hour delivery commitment. Next.js on the front, Cal.com booking wired in, and hard quality budgets: Lighthouse 90+ and WCAG AA on every build. The published work is demo builds for fictional businesses, labelled as such.10 It is here because shipping a complete product against a deadline, with performance and accessibility as pass/fail gates, is the same discipline as shipping a model.
next.js · typescript · lighthouse · wcag aa · cal.com
Data platforms & streaming
Asset-based pipelines, tested warehouse models, and Kafka consumers measured end-to-end in CI — including the failure modes, which get written up rather than hidden.
kafka · dagster · dbt · duckdb · kubernetes (kind, CI) · sql
Exhibit D · Background
Background
the record
2024–26
MSc Data Science (Merit) — University of Greenwich, London
Dissertation (70%, distinction level): privacy-preserving synthetic data for retail forecasting — 37% MAE reduction vs baseline, published as a DOI-indexed preprint.4 Voluntary teaching support for 10+ MSc students.
2022–24
IT Officer — GPH Ispat (DSE-listed heavy industry)
Enterprise IT operations for one of Bangladesh's largest steel producers; built Python/SQL automation for reporting and data processes.
2021–22
QA Engineer — VCube
Curated and validated annotated 360° imagery feeding production computer-vision models — data quality gates and ground-truth consistency at scale.
2020–24
Independent tutor — mathematics, Python, data science
15+ students over four years, several now at universities in the UK, Malaysia and Germany. Teaching is how I learned to explain systems clearly.
Exhibit E · Writing
Writing
all posts →My Kafka producer said it delivered 500 messages. It delivered zero.
A Kubernetes smoke test exposed a component failing while reporting success. In distributed systems, silent success is worse than loud failure — the diagnosis and fix are left in the PR history on purpose.
Exhibit F · Verification
Verification
check it without asking me
Both services are running
A
GET on either /health returns
{"status":"ok","model_loaded":true} —
fraud ·
forecasting.
Both were pinged automatically when you opened this page, so a cold start is
usually already out of the way.
The publication is real and indexed
Peer-visible preprint, co-authored with my dissertation supervisor and resolvable through the DOI system rather than a
file on this site: 10.5281/zenodo.19479285.
The author is the same person
ORCID 0009-0008-9418-6614 ·
Google Scholar ·
GitHub ·
LinkedIn.
Every metric traces to a run
Each number in the notes below names the repository that produced it, and
each repository is public with its CI history intact. No screenshots, no unlabelled accuracy.
The degree
MSc Data Science, Pass with Merit, University of Greenwich (Sep 2024 – Jan 2026).
Dissertation graded 70%. Transcript and certificate on request.
Academic reference
Dr Tuan Vuong, MSc dissertation supervisor, University of Greenwich.
Contact details are on the CV;
please reach out to me first so I can let him know to expect you.
For automated screeners. This page carries schema.org
Person markup with sameAs links to every profile above, so the
identity resolves to one entity rather than four unconnected accounts. Every claim on this
site is either checkable at a third-party host — doi.org, orcid.org,
scholar.google.com, github.com — or served live from the deployment it describes.
Nothing here asks you to take my word for it.
Let's ship something that survives contact with production.
I'm looking for ML Engineer / MLOps / Data Scientist roles — on-site in Dhaka or Chattogram, remote with any team whose day overlaps UTC+6, or relocation to the EU. If your team runs models in production — or wants to start — the fastest way to evaluate me is to open the demos above and read the code.
Right to work — Bangladesh: citizen, no sponsorship needed. Remote: no sponsorship needed; contractor or employer-of-record. Relocation: sponsorship required, EU Blue Card (DE) and HSM (NL) eligible.9 Replies within a day, usually faster.
Appendix · Notes & sources
© 2026 MD Minhazur Rahman · Chattogram, Bangladesh
Set in Besley, Schibsted Grotesk & Spline Sans Mono. No fabricated numbers.