Rijusmit Biswas

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.

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.

Takes the pretrained SICKLE satellite crop models (Sentinel-2, Landsat-8 and Sentinel-1 fusion) from Tamil Nadu to Andhra Pradesh and measures what transfers. I rebuilt the dataset from 1,744 farmer survey records into 509 model-ready rice fields, split by location so neighbouring farms never land on both sides, then scored three architectures (ConvLSTM, U-TAE, UNet3D) before and after fine-tuning, each against a predict-the-training-mean reference. ConvLSTM adapts best, but no fine-tuned model beats the mean reference yet, and the write-up says so.

A two-stage system that ranks news articles by how newsworthy they are, using only the text of the article. A ModernBERT encoder with LoRA adapters scores seven news values (impact, conflict, novelty, proximity, prominence, timeliness, human interest), and a ranking tower turns those into a ranked list. Trained on BBC articles and tested on a separate CNN split of 201 groups and 1,860 articles, where it beats TF-IDF + Ridge even though that baseline was trained on CNN itself.

A SpanBERT encoder with a linear-chain CRF that I wrote from scratch, used for both question answering on SQuAD v2.0 (answers as a BIO span, with unanswerable questions as a valid prediction) and named entity recognition on CoNLL-2003. The CRF loss is unit-tested against brute-force path enumeration, with optional LoRA fine-tuning and a published model on the Hugging Face Hub.

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.

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.

Forecasts daily shipment volume for each of 27 Brazilian states 14 days ahead, and models delivery-time promises, on the Olist e-commerce dataset. Backtested with a rolling origin so no test dates leak into training. LightGBM cut forecast error (WAPE) by 29.5% against a seasonal-naive baseline and beat an LSTM at most horizons. A quantile LightGBM, with the pre-registered P90 setting, gives delivery promises 2.76 days (10.9%) shorter than Olist's own at a 4.66% late rate against Olist's 4.24%.

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.

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.