Open-source tools for AI-native workflows — by Ramacharan Reddy Kasireddy.
LangModule is a collection of open-source projects focused on making AI agent infrastructure more accessible. The projects here solve real problems encountered while building production LLM systems — from document intelligence for RAG and durable state management for agentic workflows to personal knowledge management powered by AI.
Self-hosted document intelligence for RAG pipelines — the territory of LlamaParse / Reducto / Unstructured.io, running on your own stack. One call takes a PDF/DOCX/PPTX to searchable, cited, retrieval-ready chunks: layout-aware parsing with charts converted to data tables and every block traceable to a bounding box, document classification, natural chunking that never splits a table, and hybrid (dense + sparse) retrieval with reranking on six vector stores — Pinecone, Qdrant, SQLite, Postgres/pgvector, MongoDB, or Milvus, each one connection string away. PaddleOCR-VL on your GPU + Amazon Nova for judgment, at ~$0.002/page.
pip install ingestlib
Visual review UI for ingestlib — see exactly what your documents became. A first-run wizard verifies your AWS access with real calls and activates the config with zero restarts; try-before-you-commit pipeline runs; page-by-page parse review with hover-to-highlight bounding boxes; committed ingestion with live per-stage progress; a retrieval playground where every answer points to its source on the page; and lossless vector-store migration, because S3 is the source of truth. FastAPI + React.
A LangGraph checkpoint saver for Azure Cosmos DB with sync and async support. Drop-in persistence backend for LangGraph workflows with keyless DefaultAzureCredential authentication, tip-document optimization for O(1) latest-checkpoint access, and transactional batch consistency.
pip install langgraph-checkpoint-cosmos
Production-ready PostgreSQL memory for LangGraph agents. Short-term memory (checkpointer) and long-term memory (store) with one-line setup — connection pooling, lifecycle management, retry with backoff, thread cleanup, TTL auto-expiry, and optional semantic search with any embedding provider. No boilerplate.
pip install langgraph-postgres-memory
A local-first personal knowledge base implementing the LLM Wiki pattern. Feed it URLs, PDFs, YouTube videos, or audio — an LLM organizes everything into a structured, cross-referenced wiki. Query it through any AI tool via MCP (21 tools), browse in Obsidian, or search with full-text SQLite FTS5.
pip install wikinow
Leakage-safe feature-selection screening, driven by Claude Code. Point it at a dataframe and it runs an eight-stage screening — frame → viability → inventory → value integrity → partition → relevance → redundancy → verdict — and hands back a documented notebook with a verdict + reason for every column. The fold split is frozen before any per-feature target statistic runs (no silent target leakage), a shadow-permutation floor self-calibrates each feature's bar by permuting the column itself, and a corroborated leak register flags label proxies instead of flooding on well-separated data. A screening tool, not an auto-selector — it produces a ranked, reasoned shortlist; the human decides.
pip install feature-selection-playbook
AI Engineer with 3+ years of experience building production ML and GenAI systems end-to-end. MS in Computer Science from the University at Buffalo (SUNY).
Day job: Building multi-agent systems, RAG pipelines, and AI-powered automation at scale using LangGraph, AWS Bedrock, Azure OpenAI, and cloud-native infrastructure.
Core areas:
- LLM agents & orchestration (LangGraph, LangChain, MCP, ReAct, tool calling)
- RAG systems (hybrid search, embeddings, guardrails, observability)
- Cloud AI infrastructure (AWS Bedrock, Azure OpenAI, Terraform, Kubernetes)
- Deep learning (PyTorch, fine-tuning, LoRA/QLoRA, quantization, distillation)