Open-source research and tools at the intersection of multi-agent systems, mathematical foundations of AI, and interpretable machine learning.
I'm Rajesh, an AI/ML engineering professional with ~10 years in AI and ~20 years overall. My first decade was in manufacturing (automotive, aerospace, fluid machinery), and my deep interest in statistics and ML led me to data science. I've consulted with 100+ clients and built SaaS products for data/MLOps and HR platforms.
Research Focus: Multi-agent systems, mathematical foundations of AI (dynamical systems, topology, category theory), and interpretable machine learning.
- Portfolio: rajeshrs.in
- LinkedIn: linkedin.com/in/rajeshrs
Pythonic framework for building collaborative multi-agent systems. Agents coordinate through message passing via a shared "reef" (message bus), with no central orchestrator needed.
- PyPI:
pip install praval - Website: pravalagents.com
- Repo: github.com/aiexplorations/praval
- Blog: Praval Agentic AI Framework
Key features: Built-in observability (OpenTelemetry), secure spores (end-to-end encryption), storage integrations (Qdrant, MinIO, Redis, PostgreSQL), and LLM orchestration.
Fast BM25 search engine with category theory abstractions. Supports lexical, vector, and hybrid search with clean mathematical foundations.
Features vectorized search for large corpora, native HNSW indexing, and Reciprocal Rank Fusion for hybrid retrieval.
Projects investigating neural networks through the lens of dynamical systems theory.
| Project | Description | Links |
|---|---|---|
| Deep Lyapunov | Analyze neural network training stability using Lyapunov exponents and perturbation-based trajectory analysis | Repo · Blog |
| Agentic NLD | Empirical investigation of chaotic dynamics in two-agent LLM conversations with mathematical framework | Repo |
| Nonlinear Dynamics | Explorations of chaos, fractals, strange attractors, and bifurcation diagrams | Repo |
| Hamiltonian Neural Networks | Physics-based neural network optimization using Hamiltonian mechanics instead of gradient descent | Repo |
Using topological data analysis (TDA) and persistent homology to understand neural network representations.
| Project | Description | Links |
|---|---|---|
| TODACOMM | Topological Data Analysis Comparison of Multiple Models. Characterizing transformer representations using persistent homology | Repo · Blog |
| TDA-DNN | Unified pipeline for topological analysis of neural networks. H1 Paradox research | Repo |
Experimental approaches to computation and reasoning.
| Project | Description | Links |
|---|---|---|
| Vidai | Arithmetic computation via structure understanding. Separates expression parsing (learned) from computation (hardcoded). Achieves 100% accuracy on validation sets | Repo |
| Tlön Math | Process-as-primitive mathematics inspired by Borges. Processes are primitive; objects emerge as stable patterns | Repo |
Physics and optimization from my mechanical/aerospace engineering background.
| Project | Description | Links |
|---|---|---|
| Beam-Column Simulation | Timoshenko beam-column analyzer with P-Δ effects, multiple boundary conditions, and interactive FEA-style visualization | Repo |
| Genetic Algorithm Wing Optimization | Wing planform optimization using enhanced GA with adaptive mutation, elitism, and real-time visualization. Originally published in 2010 | Repo |
| Project | Description | Links |
|---|---|---|
| Praval Deep Research | Local-first AI research assistant for ArXiv papers, built with Praval | Repo |
| Praval Tutorials | Interactive Jupyter notebooks for learning the Praval multi-agent framework | Repo |
| KayGeeGo | Knowledge graph generator using Go and Neo4j with LLM-powered relationship mining and 3D visualization | Repo |
| Sangraha | Object store service with FastAPI, message queue, and MinIO | Repo |
| PulseGate | Rate limiter system implementing best practices for request throttling between services | Repo |
| EspressoInsights | LLM app observability and monitoring using Arize Phoenix with locally-run Ollama models | Repo |
| docx-mcp | MCP server for authoring Word documents with Claude and ChatGPT | Repo |
| Markdown to PPTX | Convert Markdown files to PowerPoint presentations | Repo |
| SlowDB | Educational vector database implementation for learning database internals | Repo |
| Project | Description | Links |
|---|---|---|
| GoPandas | Pandas replacement written in Go, targeting 5x+ performance over Python for common operations | Repo |
| KayGeeGo | Knowledge graph generator with Neo4j backend | Repo |
- Most projects are Python-first with emphasis on well-tested, reproducible codebases
- I organize repos by research question, bridging mathematical foundations to hands-on implementations
- Contributions welcome. Please open issues or discussions before PRs
- MIT license by default. Please cite my work if you use it in research