Projects
Pharmacokinetics · Machine Learning · Quantitative Finance
2026
PharmaSims — Mechanistic HIV ARV PK Dashboard↗ pharma-sims.vercel.app
- ▸Built interactive PK simulator modelling pH-dependent ionisation (Henderson–Hasselbalch), membrane partitioning (Shore model), and INSTI Mg²⁺ chelation across 25+ HIV antiretrovirals spanning NRTIs, NNRTIs, PIs, and INSTIs.
- ▸Derived and implemented a three-layer chelation model (ionisation × Langmuir Mg²⁺ binding × keto-enol tautomerism) explaining why raltegravir/elvitegravir underperform dolutegravir/bictegravir despite comparable Henderson–Hasselbalch fractions.
- ▸Extended to peptide/mAb electrostatics (net charge Z(pH), isoelectric point via Brent's method) for T-20 and ibalizumab, and documented cases where the passive-diffusion Shore model fails (TAF transporter override).
Mar 2026
EpiRAG — Production Agentic RAG System↗ rohanb67-epirag.hf.space
- ▸Developed a hybrid agentic RAG pipeline in Python over 19 epidemic modelling papers (~10,700 chunks), orchestrating Llama and Qwen debate agents via LangChain.
- ▸Engineered confidence-based routing (0.45 threshold) between a local ChromaDB vector store and live web search, dynamically enriched by Semantic Scholar and PubMed APIs.
- ▸Containerised and deployed the system as a Docker microservice on HuggingFace Spaces with zero-disk storage optimisation and real-time SSE debate streaming.
Dec 2025
Nifty50 Stock Forecasting System — Full Stack ML↗ nifty50-stock-predictor-dashboard.streamlit.app
- ▸Engineered a full-stack LSTM forecasting pipeline in Python, ingesting raw market data into PostgreSQL and engineering 15+ financial signals (returns, momentum, volatility).
- ▸Trained models in TensorFlow and served predictions via a highly concurrent FastAPI backend and live Streamlit dashboard.
- ▸Built a rigorous out-of-sample validation framework to systematically eliminate data leakage, overfitting, and regime instability.
Dec 2025
Adaptive PCA Portfolio Analytics↗ github.com/RohanBiswas67/Fin-pca
- ▸Built a rolling PCA factor decomposition pipeline across 9 global equity indices using adaptive 60-day windows to detect regime shifts and non-stationarity.
- ▸Implemented a full backtesting loop yielding a 2.4 in-sample Sharpe ratio, integrating drawdown analysis and regime-sensitivity diagnostics.