JSentrix

# JSentrix (C) 2025 Jasmeet Singh Bali GPL V3 # Autonomous Transaction Monitoring & Fraud Response System (Private, Multi-Agent, MCP&A2A-Compliant) ### 📌 Overview This project is a fully local, privacy-preserving, multi-agent system for real-time transaction monitoring and fraud response with approach rule based agentic graph system (AGS) built for use case with bank/fintech IT teams and can be further extended and inspired from to develop fully customizable multi agentic and multi modal flow in other domains. It leverages the **Model Context Protocol (MCP)** for modular integrations, runs all AI models and data stores locally, and features an **Electron desktop app** (for IT staff) for monitoring. ---  ## 🍿 Previews | Flow | Preview | |-------|--------------------------------------------------------------------------------------------------| | 🔑 Login| <img src="nuances/JSentrix_Login_Preview_2025-08-14.gif" alt="Login Flow" width="680" height="340"/> | | 1️⃣ A2A | <div style="display:flex-column; align-items:center; gap:10px;"> <img src="nuances/Screenshot_2025-05-20_Jsentrix_optim_pipeline_node_relationships.png" alt="Clause Knowledge Graph" width="680" hight="340" /> <img src="nuances/anim_default_mode_selection_demo.gif" alt="Default mode selection" width="550" height="250" /><img src="nuances/a2a_preview_demo.gif" alt="A2A demo preview 1" width="800"/> <img src="nuances/A2A_Mode_Demo.gif" alt="A2A demo preview 2" width="800"/></div> | | 2️⃣ A2A + MemoryCached| <div style="display:flex-column; align-items:center; gap:10px;"> <img src="nuances/a2a_cached_mode_selection_demo.gif" alt="Cached mode selection" width="550" height="250"/><img src="nuances/qdrant_memory_assesed_prior_events.png" alt="Prior assessed memory events qdrant" width="680" hight="340" /><img src="nuances/JSentrix_A2A_Caching_Mode_Demo.gif" alt="A2A cached demo preview" width="800"/></div> | | 3️⃣ Async Redis Stream | <div style="display:flex-column; align-items:center; gap:10px;"> <img src="nuances/asyncredis_mode_selection_demo.gif" alt="Async redis mode selection" width="550" height="250"/><img src="" alt="Async redis demo preview" width="800"/></div> --- ## 🏭 Architecture - **Typescript Electron App(MCP-client):** Local desktop app for IT staff, with embedded MCP client for analysis and tasking dashboard for monitoring transactions, alerts, and agent actions. - **Python MCP Server(s):** fastMCP server wrapped by fastapi that exposes tools and agents via MCP and A2A compliance agent-to-agent comms; handles transaction ingestion, agent orchestration, and local AI Triage Agent flow. - **Gateway FastAPI:** Act a minimal proxy gateway service that enables interaction between MCP Server and Electron app. - **Go+Fiber Streaming Hub with traefik setup:** Acts as centralized real time notification and event dispatcher knitted with kafka and redis fan out pub-sub pattern with multiple streaming-hub instance scalability for each client asssociated to streaming-hub get all events and behind traefik - **Local Data/Models/LLM:** All LLMs (Ollama), vector DB (Qdrant), and graph clause memory (neo4j) run locally-no cloud or API keys. --- ## 🚦 End-to-End Flow 1. **IT staff launches Electron app** on their workstation. 2. **Electron MCP client** connects to the local MCP server via localhost. 3. **User triggers transaction ingestion/analysis** from the Electron UI. 4. **MCP server** fetches transactions from a local data source (CSV, database, or direct feed). 5. **Fraud detection agent** (local LLM) analyzes transactions for suspicious patterns. 6. **If fraud is suspected:** - **A2A workflow:** - Intake Agent starts the stream for consuming txn from the target source as per end-user instructions - Assessment Agent gathers more context (prior memory events and history) along with filtering and passing the high priority transactions only to the action/agent. - Action/analysis agent infers the knowledge base with the provided context from assessment agent and llm inference and prepares freeze/escalation actions and automatically executes appropriate actions immediately for voilating txns without human intervention. - Action agent output is finally passed to Judge agent for further human in loop downstream flows along with summarization and report via summarizer model for txns that are non-conclusive by the action agent 7. **User reviews and approves/denies actions** in the Electron app. 8. **All actions, alerts, and transaction statuses** are visible in real time on Electron app dashboard. --- ## 💫 What is home-brew(in-house just used open source depend to make it alive though) !! - [x] a real time use case project how an AI system can be engineered to automate and augment human capabilities in a tx analysis system that is based on business compliance/clauses - [x] custom knowledge base prep and ingestion pipeline - [x] custom retrieval pipelines with score&decay, reranking, filter+metadata ... - [x] triage custom flow setup with multistep intake of txn streams - [x] each stream_id gets an isolated Intake → Assessment → Action agent flow, enabling per-client control, scaling, and state encapsulation - [x] minimal but effective a2a compliant agent-to-agent communication inspired by agent card and mcp compliance protocols - [x] self-improving event memory store to improve system analysis capacity the more it analyze the txn streams as time progresses to improve system for the next time it analyze the txns --- ## 💫 Core Features - **Local-Only Operation:** No third-party APIs or cloud dependencies; all data, models, and memory are local. - **MCP-Based Integration:** Modular, protocol-driven access to data sources and tools. - **Multi-Agent System:** Context- and memory-rich agents (fraud detection, investigation, notification, action) collaborating via A2A. - **Electron Desktop App:** Cross-platform, user-friendly interface for IT staff, with embedded MCP client. - **Electron Monitoring Dashboard:** Real-time visualization of transactions, alerts, and agent actions. - **Audit & Logging:** Full traceability of actions and agent decisions. - **Dockerized Deployment:** One-command setup for all components. --- ## 💡 Getting Started 1. **Clone the repo and follow setup instructions for each component in the next step.** 2. **Start the MCP server and supporting services (Ollama, neo4j) via Docker or local a/c to instructions.** ```bash # start up postgres,neo4j,qdrant,redis,kafka and dockerized streaming-hub with traefik docker instance from root jsentrix docker-compose up -d # startup ollama qwen3 model locally in terminal ollama pull qwen3:1.7b # only the first time ollama run qwen3:1.7b # makes sure venv is activated and .env is set for each of the backend components # mcp_server python -m interface.mcp_server --http # gateway (super user is auto created everytime the gateway fastapi service startsup with mcp_server health check and accessibility) uv run ./src/main.py # frontend electron app startup # mcp-client npm run start # login to mcp-client as superuser whatever email or password you set in gateway/.env make sure to restart gateway reff .example.env FIRST_SUPERUSER_EMAIL= FIRST_SUPERUSER_PASSWORD= #qwen3 http://localhost:11434 # local api qwen3 # neo4j http://localhost:7474 # qdrant ui http://localhost:6333/dashboard # gateway http://localhost:8080/docs # mcp_server http://localhost:9001 # streaming-hub (exposed via traefik with Load balancer not directly) http://localhost/health # to prep neo4j knowledge base with initial clauses manually # 💫 though mcp_server takes care of this automatically on server startup just put a sample_clause.pdf inside data dir python generate_sample.py # generate data/sample_clause.pdf python application/run_pipeline.py # generate sample_clause.md and prep and injest knowledge base neo4j # access swagger docs for running dockerized scalable streaming-hub golang fiber app http://localhost/swagger # check docker disk usage docker system df # remove all unused containers, images, networks, and build cache docker system prune ``` --- ## 🎯 JSentrix Milestones Tracks |V1 Track | Milestone | |------|----------------------------------------------------------------------------------------------- | | 1 | ~~Scaffold Electron app (React/TypeScript), set up Python MCP server, connect via localhost.~~ | | 2 | ~~Setup custom neo4j, retriever interface for both llamaindex and langchain agent support~~ | | 3 | ~~Implement transaction ingestion tool (local DB/CSV/pdf) unstructure+langchain+docling~~ | | 4 | ~~optimiz and expand retrieval interface with relationship traversal cypher utils~~ | | 5 | ~~custom flow setup including relevance decay score sort, summarization,memory aware querying~~ | | 6 | ~~add async support for retrievers, postprocessors, utils downstream pipelines~~ | | 7 | ~~sphinix doc and instrumentation with opentellemetry setup~~ | | 8 | ~~Setup reusable BaseAgent class interface with mcp+a2a compatibility for across all agent in system~~| | 9 | ~~Setup support jsonrpc2.0 for comm b/w gateway and mcp_server with rest backw compat~~ | | 10 | ~~Setup streminges and abortinges tool for mcp server and its peripheral setup with e2e websocket and streaming support with integ of kafka and redis pub/sub~~ | | 11 | ~~Intake Agent setup with message input/output integ with prior memory event, enriched txn data to be passed to next phase as list of enriched transactions to the langchain Assessment & Prioritization Agent~~ | | 12 | ~~Setup Assessment Agent and its peripherals e2e logs shud be streamed for both intake and assessment agent seprately~~ | | 13 | ~~Add A2A comm workflow i.e Assessment & Prioritization Agent comm with Analysis /Action Agent via a2a protocol internally inside the mcp_server with logs of action agent seprately streamed to ui~~ | | 14 | ~~polish Electron panels, loader, toast, split RootWrapper, vertical+horizontal split UI + streaming mode support and bug fixes + sidebar + add snapshot for v1 ui + recording for default,caching, async redisstream modes working under 1 min each mode video and add gifs and screenshots~~ | ```bash ===== JSentrix v2.0 NOTE- to research on the 🎈 Abstracted Phases for Triage Flow current approach thoroughly before moving forward here tweak and finalize then move forward with below milestones==== ``` |V2 Track | Milestone | |---------|----------------------------------------------------------------------------------------------- | | 15 | Setup Judge agent and its comm from action agent for ND txns flow check nuances/dev.md reff: 🎈 Abstracted Phases for Triage Flow section for flow along with strucutred txn output to be passed to gemini summarizer | | 16 | visual agent graph tool invocation with xy react flow payload intake-assessment-action agent logs reff RootWrapper Bottom panel XY react flow visual | | 17 | implement RAG HyDE (Hypothetical Document Embeddings) strategy approach ideally at the Intake level need to reserach and brainstorm it more in reff to the e2e flow | --- ## ⚙️ Tech Stack - **Electron+Typescript** (React/TypeScript) - Desktop app & MCP client - **Python+FastMCP+FastAPI** - MCP server, agent orchestration via Langchain - **Golang+Fiber+Traefik** - Fiber streaming-hub multiple replicas with traefik as proxy server and load balancer - **Ollama** - Local LLMs for analysis(qwen3:1.7b) + summary(Gemma3:1b) - **Langchain + Llamaindex** - for Triage Self improving custom agent setup with MCP and a2a compliance and custom agent orchestration and jsentrix custom flow setup with custom retrievers, score and decay patterns, reranking etc... - **neo4j** - Local graph database for knowledge base clauses - **postgresql** - Client/Electron user login both admin/non-admin users via gateway - **qdrant** - store past inferenced and processed events stored in memory for future txn processing in triage self improving jsentrix flow - **redis** - pub/sub channel broadcasting internal streaming-hub with FAN out pattern sending single consumed message by kafka partition to other streaming-hub replica subscribers of process,events by mcp_server and its agents to client electron ui realtime events streaming via websocket - **kafka** - single/dedicated streaming-hub replica for consuming events published/forwarded by mcp_server agents and processes - **Docker** - Deployment and local orchestration - **Sphinix** - Docs - **OpenTelemetry** Tracing and Instrumentation --- ## 🔃 Tech/Tools/Framework Doc References - [MCP Protocol](https://github.com/anthropics/mcp) - [Ollama](https://ollama.com/) - [Langchain](https://python.langchain.com/docs/introduction/) - [LlamaIndex](https://docs.llamaindex.ai/en/stable/#introduction) - [Neo4j](https://neo4j.com/docs/operations-manual/current/docker/introduction/) - [Qdrant](https://qdrant.tech/documentation/) - [Electron](https://www.electronjs.org/) - [Sphinix](https://www.sphinx-doc.org/en/master/usage/installation.html#pypi-package) - [OpenTelemetry](https://opentelemetry.io/docs/languages/python/) --- ## License [](https://choosealicense.com/licenses/gpl-3.0/)