# ⚡ Nexora AI > **Portfolio Case Study & Technical Deep Dive** > *An enterprise-grade autonomous multi-agent AI platform uniting dynamic intent routing, multimodal RAG, real-time generative UI, and defense-in-depth security.* ``` • Live Repository : https://github.com/Sandeep-singh-99/Nexora • Architecture : Asynchronous Multi-Agent Graph (LangGraph) + Real-Time SSE Gateway (FastAPI) + Modern Web (Next.js 16) • Core Domains : Autonomous Agent Routing, In-Memory RAG, Generative UI, Enterprise Security ``` --- ## 💡 The Problem & The Nexora AI Solution - **The Problem:** Conventional LLM applications rely on monolithic prompts that hallucinate across complex tasks, output static plain markdown, store sensitive user files directly on server disks, and waste inference tokens processing malicious jailbreaks. - **The Solution:** **Nexora AI** decouples cognition into a stateful **LangGraph multi-agent network**, replaces static text with **interactive Generative UI** over SSE, enforces **zero-disk in-memory RAG** with Human-in-the-Loop PII interception, and prevents prompt injection pre-inference. --- ## 🏛️ System & Multi-Agent Architecture ```mermaid flowchart TD Client["Next.js 16 Client\n(React 19 / Tailwind v4 / Motion)"] <-->|HTTP & Real-Time SSE Stream| API["FastAPI Gateway\n(JWT Auth & Double-Submit CSRF)"] API --> InGuard["Pre-Execution Guardrail\n(Length, Abuse & Injection Filter)"] InGuard --> Router["LangGraph Intent Classifier & Router"] Router --> Chat["General Chat & Web Agent\n(Tavily Real-Time Grounding)"] Router --> Code["Coding Specialist Agent\n(Software Architecture & Logic)"] Router --> Math["Symbolic Math Engine\n(SymPy & SciPy Deterministic Run)"] Chat & Code & Math --> OutGuard["Output Sanitizer\n(Credential & Secret Redaction)"] OutGuard --> API Chat <--> PGVector[("PostgreSQL + pgvector\n(768-dim HNSW Index)")] DocIngest["In-Memory Parser (PDF/DOCX/OCR) & YouTube Ingester"] --> PGVector ``` --- ## ✨ Core Features & Technical Highlights - 🤖 **Autonomous Multi-Agent Routing:** Classifies query intent via structured output to dispatch to specialized Chat, Code, or Math agents with zero manual prompting. - 📄 **Zero-Disk In-Memory Document RAG:** Streams PDFs and DOCX files into memory buffers (`io.BytesIO`) with OCR fallback and PII detection—no data ever touches the disk. - 📺 **Interactive YouTube Video RAG:** Ingests video transcripts into timestamped semantic windows; interactive player seeks playback to the exact second. - 🎨 **Dynamic Generative UI Registry:** Emits structured payloads over SSE to hydrate rich widgets (video players, clocks, math derivations, charts) directly in chat. - 🧮 **Hybrid Symbolic Math Engine:** Deterministically computes calculus, algebra, and matrices via SymPy/SciPy, rendered with KaTeX step-by-step proofs. - 🛡️ **Dual-Stage Security Guardrails:** Intercepts jailbreaks and prompt injections before LLM invocation (0 token waste); redacts secrets and credentials from outputs. - 🏷️ **Context-Aware Title Generation:** Synthesizes concise 3–5 word conversation titles dynamically as discussions unfold. - 📌 **Pinned Messages & Thread Memory:** Pin key answers for immediate reference; persist session states across turns with LangGraph checkpointers. - 🌐 **Real-Time Web Search Grounding:** Integrates live web search via Tavily for verified, up-to-date answers accompanied by publisher citation cards. - 🔭 **End-to-End Observability:** Full distributed tracing of agent nodes, tool latency, token consumption, and state transitions via LangSmith. --- ## 🎨 Dynamic Generative UI Registry Nexora AI replaces plain text responses with interactive, client-hydrated UI widgets: | Feature Area | Interactive Component | Visual & Functional Capability | | :--- | :--- | :--- | | **YouTube Video RAG** | `` | Synchronized video player with searchable captions and click-to-seek timestamps. | | **Symbolic Math** | `` | Deterministic calculations with step-by-step derivation and KaTeX formula rendering. | | **World Clock** | `` | Live regional time display with timezone offsets and localized country flags. | | **Web Citations** | `` | Live fact attribution cards linking directly to external sources with favicons. | | **Data Visualizer** | `` | Dynamically rendered analytical charts (bar, line, distribution) from structured data. | | **Data Grid** | `` | High-density tabular grid with column sorting, alignment, and markdown support. | | **Engineering Plan** | `` | Multi-phase architecture blueprints and task breakdown progress cards. | | **Executive Callout** | `` | Highlighted cards summarizing key definitions, caveats, or strategic takeaways. | --- ## 💡 Key Engineering Challenges Solved 1. **Eliminating Mathematical Hallucinations:** Offloaded calculations from probabilistic LLM token prediction to deterministic Python `SymPy` and `SciPy` engines with KaTeX proofs. 2. **Enterprise Privacy & Zero-Disk Storage:** Engineered an in-memory RAG pipeline using `io.BytesIO` streams, automatically purging document buffers after pgvector vectorization. 3. **Flicker-Free Streaming UI:** Implemented a typed SSE protocol separating `token`, `ui`, and `citation` events to atomically render React components without stream interruption. 4. **Zero-Token Threat Neutralization:** Built pre-execution regex and rule-based guardrails that reject jailbreak attempts and prompt injections before calling the LLM.

Nexora AI

Built with