Debopriyo — Data scientist

Portfolio · Data science · Rev. 2026Open to new opportunities

Fig. 1Particles initialised as Gaussian noise and fitted to the target glyphs by gradient descent with learning-rate warm-up and annealed noise. Move your cursor through the figure to perturb the data — or .

log lossepoch

epoch: 0000
loss: —

Data scientist — I turn noisy data into models people can trust, and decisions they can act on.

Debopriyo^1,2

¹Self-taught, 4+ years in industry & open source

²M.S. Computer Science, in progress

↓ Read the experiments ✉ Correspondence

§1

I’m a data scientist who learned the craft by building: four-plus years of hands-on work, now formalised through a Master’s in Computer Science.¹

I work across the whole arc of a data problem — framing the right question, wrangling messy sources, engineering features, training and evaluating models, and shipping the result as something people actually use. I care as much about the error bars as the headline number.²

Keywords —Machine learning · Statistical modelling · Deep learning · NLP · Data engineering · MLOps · Visualisation

§2

Selected work, pulled live from GitHub. Each row is a repository — open it for code, data and notes.

Table 1Selected experiments. Click a column header to sort.

Topics

E-01

visual symphony

Bring images to life with sound! This Python tool uses AI to create audio narratives from images via VisualSymphony.

Python

Python—1

E-02

Drowsiness Detection

Notes and code on GitHub.

Python

Python—2

E-03

ocean trash

Notes and code on GitHub.

Jupyter Notebook

Jupyter Notebook—1

E-04

sgay

🏆 Finalist Project - Team 01 GT | codesummit@sikkim'50 Hackathon

HTML

HTML—1

E-05

lumen

Notes and code on GitHub.

TypeScript

TypeScript—1

Source: GitHub API, refreshed hourly.Full record → github.com/Dragon4926

§3

The same loop underpins every project, whether it ends in a paper, a dashboard or a production API.

↺ retrain on drift

01Question
Pin down the decision the analysis has to inform — and the metric that proves it.
02Data
Source, join and audit. Most of the signal (and most of the bugs) live here.
03Features
Encode domain knowledge. Simple, well-understood features beat clever ones.
04Model
Start with a baseline, earn every bit of added complexity.
05Evaluate
Honest validation, error analysis, uncertainty — not just a single score.
06Deploy
Package as an API, a pipeline or a dashboard that people actually open.
07Monitor
Watch for drift, close the loop, retrain when the world moves.

01Question
Pin down the decision the analysis has to inform — and the metric that proves it.
02Data
Source, join and audit. Most of the signal (and most of the bugs) live here.
03Features
Encode domain knowledge. Simple, well-understood features beat clever ones.
04Model
Start with a baseline, earn every bit of added complexity.
05Evaluate
Honest validation, error analysis, uncertainty — not just a single score.
06Deploy
Package as an API, a pipeline or a dashboard that people actually open.
07Monitor
Watch for drift, close the loop, retrain when the world moves.
↺ Monitor feeds back into Data when the world drifts.

Fig. 2Working method as a directed graph. Dashes indicate the direction of data flow; the red edge is the retraining loop.

§4

Not a list of logos — where each tool actually sits in the workflow. Hover a cell to trace its row and column.

ColCollectClnCleanModModelEvalEvaluateDepDeploy

Python

SQL

pandas

NumPy

scikit-learn

PyTorch

TensorFlow

Matplotlib

PostgreSQL

FastAPI

Docker

AWS

Fig. 3Tools by pipeline stage. primary occasional

§5

A live cell. Type an expression and press Enter — try me.skills or help().

portfolio.ipynbPython 3

In [1]:

from portfolio import Debopriyo
me = Debopriyo()
me.summary()

Out[1]:

	field	value
0	name	Debopriyo
1	role	Data scientist
2	experience_yrs	4+
3	education	M.S. CS (in progress)
4	focus	applied ML, modelling, data products
5	status	open to opportunities

6 rows × 2 columns

In [2]:

Enter to run · Tab to complete · ↑/↓ for history