Pipelines that carry data, and models that read it.
I build the batch and streaming infrastructure that moves data at scale —
Spark, Airflow, dbt, Kafka, Snowflake — and co-author applied machine
learning research across green computing, industrial prognostics, financial
risk and medical imaging.
Data infrastructure on one side, applied machine learning research on the other.
They look separate. They are not — most of what goes wrong in a published model
is a data problem, and every repository below was built on that observation.
⏱
Ingest it
Batch and streaming pipelines over structured and unstructured sources.
Spark, Airflow, dbt, Kafka. Late-arriving events, replay after an outage,
and producers that change schema without telling anyone.
◧
Model it
Warehouse and lakehouse design — staging, conformed dimensions, slowly
changing dimensions, tested marts. Snowflake and Azure Synapse in production;
DuckDB when a model should build in CI without credentials.
🛡
Guard it
Data contracts that fail the build, column-level lineage, and PII
classification that follows the data downstream instead of going stale
three months after somebody wrote it down.
Research implementations
Papers, made runnable
Reference implementations of methods from papers I have co-authored. Each ships an
evaluation protocol that does not cheat and tests that fail when it does.
None of them quote an accuracy figure — they ship no dataset, so any number
would be a re-run on different data or a copy from a paper this code did not produce.
Both are ways of being wrong in public.
green-anomaly-detection · measured
The one exception, because energy accounting is the contribution:
every detector reports its joules and grams of CO₂e beside its detection score.
Produced by scripts/benchmark.py, three seeds, 6000 flows.
Detector
Tier
Avg precision
Joules
AP / joule
mahalanobis
cheap
0.231
0.14
1.62
z-score
cheap
0.147
0.07
2.29
one-class-svm
expensive
0.182
9.77
0.019
isolation-forest-32
moderate
0.132
8.54
0.016
lof
expensive
0.296
38.50
0.008
isolation-forest
moderate
0.132
25.01
0.005
elliptic-envelope
moderate
0.192
148.54
0.001
Mahalanobis reaches 0.231 average precision for 0.14 joules. LOF reaches 0.296
for 38.50 joules — buying those last 0.065 points costs 275× the energy.
Whether that trade is worth taking depends on your alert budget; the point is that
you can now see the trade at all.
Oncology engineering
Cancer, at the scale it actually arrives
Engineering projects rather than paper reproductions — built to be read, run
and checked. Each one exists because a specific thing in cancer machine learning is
routinely got wrong, and each measures how much that thing is worth.
Clinical data and diagnostics
Healthcare, including the unglamorous half
Two of these are clinical data infrastructure — DICOM ingestion and PHI
de-identification — the part of healthcare AI that consumes most of the effort
and gets written up least. Three are diagnostic models, each chosen because the task
is scored wrongly in the published literature in a specific, measurable way.
Data engineering
The infrastructure half
The patterns I reach for in production, written as libraries with the reasoning in
the docstrings and a test for every claim. All four run their suites in CI on
Python 3.10, 3.11 and 3.12 — no cluster, no credentials.
Education
Background
2023 — 2025 · CONNECTICUT
MS, Computer Science
University of New Haven
2017 — 2022 · DHAKA
BSc, Computer Science & Engineering
BRAC University
Undergraduate thesis on voice impersonation detection using LSTM-based
recurrent networks with LIME explanations, supervised by Dr Md Golam Rabiul
Alam.