Bireswar Sett

Seven years of programming, from QBasic to bare metal.

I’m Bireswar Sett, a Data Science student at IIT Madras and a Computer Science student at SOA University. I build systems software, algorithms, and machine learning, and I work from first principles.

See selected work Download résumé

How the last seven years went

  1. 2018QBasic

    First programs.

  2. 2019Java

    Serious programming begins.

  3. 2020C# and Unity

    Interactive, real-time software.

  4. 2024Two degrees

    BS Data Science at IIT Madras and B.Tech CSE at SOA, in parallel.

  5. 2025Bare-metal systems

    StarOS: an x86 bootloader and kernel for physics simulation.

  6. 2026Research and applied AI

    k-centre optimization, a custom MCQ model, and Akasha.

Selected work

Three projects across applied AI, operating systems, and machine learning.

Akasha

Team lead, six engineers

Aug 2026 – Sep 2026

Source on GitHub

An agentic vision-language platform that answers natural-language questions about optical, SAR, and multi-temporal satellite imagery. Built against an ISRO problem statement and delivered as a deployment-ready system.

What I built

  • Query understanding, task routing, and specialist geospatial model invocation.
  • Multimodal evidence synthesis, so every answer traces back to model outputs.
  • Model-serving infrastructure, backend APIs, authentication, cloud storage, and containerized services.
  • Led a cross-functional team of six through design, implementation, and deployment.

Stack Python, FastAPI, React, Firebase, Google Cloud Run, Hugging Face, Docker

StarOS

Independent project

Oct 2025 – present

Source on GitHub

A bare-metal x86 operating system for real-time N-body physics simulation. It is a study in kernel engineering and in the foundations deterministic simulation needs.

Status

  • Done: x86 bootloader and the 16-bit real mode to 32-bit protected mode transition.
  • In design: memory management, interrupt handling, hardware initialization, and a custom scheduler for simulation workloads.

Stack C, C++, x86 Assembly

Smart MCQ Solver

Machine learning project

Jul 2026 – Aug 2026

Source on GitHub

A Transformer encoder written for multiple-choice question answering. Shared embeddings and cosine-similarity ranking score each candidate answer.

What I did

  • Benchmarked the model against BERT and RoBERTa.
  • Tuned hyperparameters and tracked experiments with Weights & Biases.
  • Reached about 0.75 Mean Average Precision on the public Kaggle dataset.

That score comes from my own evaluation and has not been independently verified.

Stack Python, PyTorch, Hugging Face, Weights & Biases

Independent research

Algorithm design, driven by one question: how much structure is hiding in a hard optimization problem?

A heuristic for the discrete Euclidean k-centre problem

Independent research, Feb 2026 – present

The goal is to minimize the worst-case covering radius for k centres chosen from the input points in 2D. The heuristic partitions points with K-means, initializes each cluster with its discrete 1-centre, then refines the result with a deterministic one-centre-swap local search.

I’m now studying optimal-centre configurations, adjacency constraints in one dimension, and how the structure generalizes to higher dimensions.

Internal results. Not published research.

These numbers come from my own internal testing. They have not been peer-reviewed, published, or independently verified. Please read them as preliminary engineering measurements, not as established findings.

Internal benchmark against exact optima
MeasureThis heuristicGonzalez + local search
Exact recovery71.9% (302/420)65.0% (273/420)
Mean approximation ratio1.017×1.031×
Worst-case ratio1.301×1.414×

Setup: 420 synthetic instances of 120 points each, drawn from 12 distribution families. Lower ratios are better. These are empirical results on synthetic data only, and no theoretical guarantee is claimed.

Background and skills

A broad computer science foundation, strongest where software meets hardware and mathematics.

I wrote my first program in QBasic in 2018, moved to Java in 2019, and have kept building since. My work now spans backend infrastructure, low-level systems, GPU computing, algorithm design, and machine learning.

I like to understand a problem from first principles, make deliberate engineering choices, and carry an implementation all the way to a working system. I also care about clear technical writing and taking ownership of complex work.

Education

B.S., Data Science and Applications
Indian Institute of Technology MadrasJul 2024 – Sep 2028
B.Tech, Computer Science and Engineering
Siksha ‘O’ Anusandhan UniversityAug 2024 – Aug 2028
Languages
C++, C, Python, Java (since 2019), C# (since 2020), JavaScript, SQL, x86 Assembly, Bash
Systems and GPU
x86 Assembly, C, C++, CUDA, Linux, Bash
Machine learning
PyTorch, scikit-learn, Hugging Face Transformers, NumPy, Pandas, Matplotlib
Backend and data
Flask, Django, PostgreSQL, SQLite
Tools
Git, GitHub, Vim, VS Code, Visual Studio