Student @ IIT Ropar | singhpranav431@gmail.com
I build AI systems at the intersection of research and production. My work spans probabilistic machine learning, LLM systems, scientific ML, and applied full-stack AI products.
Current focus areas:
- Probabilistic memory systems for autonomous AI agents
- Physics-informed neural networks and uncertainty quantification
- Scalable LLM pipelines and MLOps
nous-state pip install nous-state
A probabilistic memory engine for long-running AI agents. Instead of storing facts, it maintains Bayesian belief distributions over entity attributes and resolves contradictions mathematically using surprise-driven updates and immutable delta logs.
- Zero runtime dependencies (pure Python stdlib)
- Surprise scoring via Shannon self-information:
S = -log2(P(observed | model)) - Principled forgetting via entropy decay
- Full audit trail with time-travel queries
Built because vector databases have no principled way to handle belief revision when an agent's knowledge of a user changes over time.
| Project | Description | Stack |
|---|---|---|
| NyayaSahayak | AI legal assistant for Indian law | LLMs, FastAPI, RAG |
| Physics-Informed Neural Networks | Inverse modeling with passivity constraints and uncertainty quantification | PyTorch, JAX |
| Replai | AI conversation assistant | FastAPI, React, Supabase |
| Multi-Agent Job Screening | LLM agent pipeline with embedding-based candidate ranking | LangChain, Embeddings |
| Hybrid LSTM-GARCH Option Pricing | Quantitative finance model combining deep learning and volatility modeling | PyTorch, statsmodels |
ML / AI: PyTorch, TensorFlow, scikit-learn, LLM APIs, RAG, PINNs, Bayesian inference
Systems: Python, FastAPI, Flask, SQLite, Docker, REST APIs
Frontend: React, JavaScript
Infrastructure: AWS, GCP, Linux, Git, MLOps
Other: C, C++, MATLAB, OpenCV, MongoDB, MySQL