Portfolio
Abtahi BinJahangirChowdhury
Data Analyst · ML Engineer
I turn raw, messy data into business-ready insight — Python and PostgreSQL cleaning, SQL analysis and reconciliation, and Power BI and Retool dashboards that answer a real business question.
About
Analysis that ends in a decision, not a chart.
I'm a Computer Science graduate from BRAC University who ended up on the analyst side of the stack. Most of my work starts in Python or PostgreSQL profiling a raw table, deciding what actually counts as a genuine record, and then answering business questions by analysing the data using SQL and Python — and ends in a Power BI or Retool dashboard someone can read without me in the room.
The AI/ML work runs alongside it rather than apart from it: a multi-task U-Net that segments and classifies breast ultrasound scans in one pass, and an undergraduate thesis benchmarking nine LLMs on Bangla summarization.
What carries across all of it is the order of operations, not the domain: pin down the question precisely, get the data honest before trusting it, and let the answer stand only as far as the evidence actually reaches. The question itself changes a lot: which rental segments are worth signing, why some patient visits never reached the billing system, whether a tumour is benign or malignant, whether a metric is even measuring what it claims to. That sequence doesn't.
By the numbers
6,612
EHR encounters reconciled against a billing database in the healthcare RCM case study
98,666
orders modelled end-to-end in the Olist Power BI pipeline
30,234 → 5,280
raw NYC Airbnb listings cleaned down to genuine short-term rentals
9
LLMs benchmarked in the Bangla summarization thesis
55+
problems solved on LeetCode
Selected Work
Six projects, each built end to end.
From raw CSV to a dashboard a stakeholder can act on, from an EHR export reconciled against a billing database, and from an ultrasound dataset to a model that segments and classifies in a single pass. Every card links to its repository.
- Data Analytics
Healthcare Revenue Cycle Data Operations Case Study — open the repository on GitHub
Four revenue-cycle problems for a healthcare billing company. Reconciled 6,612 EHR encounters against the billing database with pandas and a PostgreSQL anti-join, traced Medicare payment gaps through the CMS fee schedule (GPCI, multiple-procedure reduction), and built a 4-tab Retool dashboard on PostgreSQL.
282 of 6,612
Encounters (4.26%) never imported — an import filter blocked visits carrying non-CPT flag codes, missing 98.6% of the time vs 1.2% for others
- Python
- pandas
- PostgreSQL
- SQL
- Retool
- Data Analytics
Olist E-Commerce Sales Analytics — open the repository on GitHub
End-to-end pipeline on 98,666 Brazilian marketplace orders: raw CSVs loaded into PostgreSQL, reshaped into a reporting layer of SQL views, and modelled in Power BI as a nine-page dashboard on a DirectQuery composite model with 25 DAX measures.
2.3 vs 4.2
Average review score for late-delivered orders versus on-time — late delivery, not the product, drives the negative tail
- PostgreSQL
- SQL
- Power BI
- DAX
- DirectQuery
- Data Analytics
NYC Airbnb Market Analysis — open the repository on GitHub
A SQL → Python → Power BI workflow answering which NYC neighbourhood × room-type segments a property management company should prioritise. Cleaned 30,234 raw listings down to 5,280 genuine short-term rentals, then answered five business questions with CTEs, window functions and CORR().
$470/night
Manhattan hotel rooms — the dataset's highest median price, and simultaneously its lowest-demand segment in every borough
- PostgreSQL
- Python
- Power BI
- DAX
- AI / Machine Learning
Breast Ultrasound Segmentation & Classification — open the repository on GitHub
A multi-task deep learning system that localises a tumour and calls it benign or malignant in a single pass. One U-Net encoder is shared between a pixel-wise segmentation decoder and an image-level classification head, trained on the BUSI ultrasound dataset.
2 heads, 1 encoder
Combined Dice + binary cross-entropy segmentation loss and categorical cross-entropy classification loss, trained jointly as a weighted sum
- Python
- TensorFlow
- U-Net
- OpenCV
- pandas
- AI / Machine Learning
Vehicle Fuel Consumption Prediction — open the repository on GitHub
A three-class classifier sorting vehicles into low, medium and high annual fuel consumption from 34 years of US EPA fuel-economy data — cleaned from 38,113 raw records, class-balanced to 8,046, then benchmarked across KNN, SVM and Gaussian Naive Bayes.
97.9%
Test accuracy from the winning model, KNN (k = 3), at an F1-macro of 0.98 across all three classes
- Python
- pandas
- Scikit-learn
- Graphics
Velocity Run — open the repository on GitHub
A 3D endless runner built on raw OpenGL transformations and hierarchical modelling — every object composed from cuboids, spheres and cylinders. A procedurally generated three-lane corridor recycles its floor tiles for an infinite track, while speed and spawn rate scale up the longer you survive.
First ↔ third person
Switchable camera, dynamic rotating-bar obstacles, a projectile system and timed power-up gems
- Python
- PyOpenGL
- 3D Graphics
Research Spotlight
Exploring Cross-Domain Bangla Text Summarization Using LLMs
B.Sc. Thesis · BRAC University · January 2026
Evaluating nine large language models on zero-shot Bangla summarization across two very different domains — and finding that the metrics the field relies on quietly disagree with the summaries humans judge as faithful.
01
Role
Owned data collection and preprocessing within a five-member NLP research team, spanning two domains: Prothom Alo news articles and the XL-Sum benchmark.
02
Evaluation framework
Co-built a multi-layered evaluation stack — ROUGE and BLEU for lexical overlap, a blind LLM-as-a-Judge protocol, SBERT semantic similarity, and a hand-built error taxonomy.
03
Finding
Lexical metrics scored near zero on summaries that were in fact faithful, exposing how poorly n-gram overlap transfers to Bangla. Encoder–decoder models BanglaT5 and mT5 outperformed the decoder-only LLMs.
- Python
- Hugging Face Transformers
- SBERT
- LLM APIs
- ROUGE / BLEU
Capabilities
The stack behind the work.
Grouped the way they're actually used — the database and BI layer most of the analytics work runs through, the Python and ML side, and the habits that hold a solo project together.
Technical
Databases & BI
- PostgreSQL
- SQL (CTEs, joins, window functions)
- Power BI (DAX, DirectQuery)
- Retool
Programming
- Python (pandas, NumPy, Scikit-learn)
- C / C++ (Basic)
Machine Learning & AI
- PyTorch
- TensorFlow
- Hugging Face Transformers
- Deep Learning
- Computer Vision
- NLP
- Matplotlib
Tools & Version Control
- Git / GitHub
- VS Code
- Google Colab
- LaTeX
- Excel
- Canva
Operating Systems
- Windows
- Linux
Problem Solving
- 55+ problems solved on LeetCode
Soft skills
Collaboration & Communication
- Team collaboration
- Data storytelling
- Emotional intelligence
Work Approach
- Analytical & critical thinking
- Attention to detail
- Time management
- Self-directed learning
Experience & Education
Where the practice comes from.
Academic Research & Independent Projects
Led data collection and processing for a 5-member NLP research team (Bangla text summarization using LLMs).
Prepared analytical reports and data summaries directly for individuals' own datasets, outside any organization.
Designed promotional materials for product showcases on request.
Study forex technical analysis; perform daily price-action analysis on the XAU/USD (gold) chart.
Managed multiple independent requests and deadlines on my own, without a team or employer overseeing the work.
Education
BRAC University
Bachelor of Science in Computer Science
CGPA 3.5 / 4.00
Dhaka City College
Higher Secondary Certificate (HSC)
GPA 4.67 / 5.00
Ideal School and College
Secondary School Certificate (SSC)
GPA 5.00 / 5.00