Rijusmit Biswas
Portrait of Rijusmit Biswas

Rijusmit Biswas

Software and Machine Learning Engineer, Data Scientist, Researcher. Equipped with agentic AI tools — Claude, LangChain, LoRA, QLoRA, Pinecone, HuggingFace.

New Delhi, India.

I build software end to end and stay with it once it goes live. In practice that means the React screens, the API in the middle, the database underneath, and the servers it all runs on, plus some machine learning when a problem calls for it.

12+Live deployments
1First author papers
150+LeetCode problems
1403Codeforces rating
13Models evaluated

Head of Product Development at Student Council, IIIT Delhi

A student services portal used across the institute

  • Built and shipped a portal that more than 500 students use, with React on the front and Node.js, Express, and PostgreSQL behind it.
  • Own the database, the indexing, and the API the rest of the team builds against.
  • Set up continuous integration with automated tests, and review the code for a team of eight.
  • Swapped manual data entry for forms that validate, which cut the admin work by about 60 percent and put a stop to a recurring class of bad records.
Product engineeringVisit

Software Engineering Intern at Shipzee

A courier aggregation product, full stack work and a live stack migration

  • Shipped features across a live courier aggregation product, from the React screens through the API down to the database, working with people inside the company and with outside partners.
  • Led a move to a new stack so the system would scale and stay maintainable, and planned the switch so no one saw an outage.
  • Chased down production failures that crossed service boundaries using logs and request traces, and wrote tests around the outside integrations that caught mismatched contracts before release.
Engineering internshipVisit

AI Engineer and Undergraduate Researcher at MIDAS Lab, IIIT Delhi

An evaluation tool for checking model output quality

  • Built an evaluation tool that scores model output against several notions of alignment and flags on its own when quality drops. The checking step cut made up answers by about 20 percent, and the lab still uses it.
  • Ran the pipelines over three datasets so a rerun gives the same result, with retries and rate limit handling for outside APIs.
  • Trained models with LoRA and QLoRA at four bit precision to fit the memory we had.
  • First author on a paper about the work, now under review at AAAI 2027, with the harness and benchmarks released as open source.
Research

Machine Learning Research Intern at Computational Social Science Lab, IIIT Delhi

A satellite imagery pipeline for crop prediction

  • Built a satellite imagery pipeline on Google Earth Engine with preprocessing that runs in parallel, producing segmentation masks and inputs with several channels at the scale of a district.
  • Compared several CNN backbones on predicting phenology dates, sorting crops by type, and estimating yield, all under one shared setup for a fair comparison.
  • Measured how far the data drifts across more than ten districts using difference in differences and fixed effects models, so accuracy was checked outside the training area rather than on one held out split.
Research internshipVisit

Data Science Intern at Complex Systems Lab, IIIT Delhi

Messy real world classification and interactive visualisation

  • Led a team of five building a 25 class classifier on messy real world data, reaching 84.5 and 74.5 percent on two splits, and traced the gap to a shift in the data between them.
  • Built web apps that visualised complex datasets in real time, pulling in open data like OpenFoodFacts, RecipeDB, Carbon Footprint DB, and FNDDS.
  • Ran code reviews and mentored the junior members.
Data internship

Open Source Contributor at Meshery, Layer5

Layer5 documentation and Meshery.io

  • Sent bug reports, feature requests, documentation fixes, and pull requests.
  • Took part in code reviews and gave feedback aimed at keeping the code easy to maintain.
Open source
01

Executive Career Agent

Agent swarm / Desktop app

A private, local-first workspace that helps a senior executive run a deliberate job search. Seven cooperating agents (scout, reader, curator, assessor, scribe, conductor, harvester) discover, read, de-duplicate, score, and tailor roles against the candidate's real, self-declared experience. Runs fully offline on a bundled corpus or against the real internet. Features anti-fabrication validator that checks every claim against the Master Résumé, deterministic fit scoring via rule-based rubric, and MCP server for Claude Desktop integration.

ReactTypeScriptElectronNode.jsMCPLangChainDocker
Notes
02

TURBO

ML research

My own reimplementation of the paper by Goel and colleagues (arXiv 2502.07391). It fuses image features from a Vision Transformer with a modified BART text encoder, trains on the MORE+ dataset, and is scored with BLEU, ROUGE, and METEOR.

PyTorchViTBARTHugging Face
Notes
03

MineMEETS

Multimodal retrieval

Pipelines that turn images, text, and audio into embeddings behind one interface. Every step checks its dimensions and is safe to run again, so a failed run just gets rerun. Pinecone namespaces keep reindexing and rollback safe, and the test suite plus formatting and type checks run on every pull request.

CLIPWhisperPineconeDockerPytest
Notes
04

Flourish

Agentic service

A plant care assistant. The FastAPI backend runs several language model tool calls in sequence against outside APIs, with Firebase handling sign in and storage. Retries and graceful fallbacks keep a failure upstream from breaking what the user is doing.

FastAPILangChainFirebaseDocker
Live
05

Cross-Portfolio Default Risk Prediction

Data science research

Dual-study data science investigation into credit default prediction across credit cards (UCI, 30K rows) and installment loans (LendingClub, 5K rows). Engineered 31+ features including utilization ratios, payment trends, delinquency aggregates, and sorting-smoothing probability of default estimates. Built 13 models with isotonic calibration, SHAP explainability, cost-sensitive threshold optimization, and business impact assessment. XGBoost achieved 0.7769 ROC-AUC on cards; Random Forest 0.7467 on loans. Estimated $18.7M annual savings (18,641% ROI) on card portfolio.

PythonXGBoostLightGBMRandom ForestSHAPscikit-learnPandasOptuna
Notes
06

DreamDOT

Distributed backend

A platform for creators with subscriptions, pay per view, and media sharing. The backend is a set of GraphQL services that answer queries spanning several of them without firing a cascade of extra requests. I set the service boundaries and the schema contracts so teams could work in parallel without breaking the shared graph.

GraphQLNode.jsMicroservicesMongoDBDocker
Code
07

Hostiggo

Production SaaS

I took this from an idea to a product running on AWS. I own the React front end, the whole backend, the database, and the servers, and I put continuous integration and monitoring in place from the first week so releases stayed calm as more people used it.

Node.jsReactPostgreSQLAWSDocker
Code
08

Entropy

Web application

A study space with an assistant that helps you work through doubts. Python and FastAPI power the async endpoints that generate personalised quizzes and study material through language model adapters you can swap out.

Next.jsFastAPILangChainDockerPostgresPrisma
Live
Skills

Languages

Python, TypeScript, JavaScript, C++, SQL, C

Backend and APIs

Node.js, Express, FastAPI, REST and GraphQL, GraphQL federation, WebSockets, retries and rate limiting

Frontend

React, Next.js, Tailwind CSS, dashboards, live interfaces over WebSockets

Infrastructure and delivery

AWS, Docker, GitHub Actions, Linux, Vercel, Pytest, structured logging, static analysis

Data and distributed systems

PostgreSQL, MongoDB, Prisma, Pinecone, schema design, indexing, PySpark, Airflow, ETL

Machine learning and research

PyTorch, Hugging Face, LoRA and QLoRA, LangChain, RAG, CLIP and Whisper, Google Earth Engine

Agentic Development

Claude, Kiro, Copilot, Codex, Ollama

Publication

Publication one, Under review at AAAI 2027

Title undisclosed, revealed after review

My first author work, with A. G. Kumar, A. Anand, E. Cambria, Z. Wang, and R. R. Shah. It is a tool that scores what a model generates against several notions of alignment and flags on its own when the quality drops. The checking step cut made up answers by about 20 percent. The harness and benchmarks are released as open source.

Contact

Let's build something useful.

If you have a hard systems problem, a project worth doing, or a good research lead, email is the quickest way to reach me.

Open to roles: SDE, SDE Intern, MLE, MLE Intern, Data Scientist, DS Intern, Data Engineer, Data Engineering Intern.

rijusmit.biswas@gmail.com
Rijusmit Biswas, New DelhiKept current through 2026