MCP tool server
Calendar, notes, vision, memory and autonomous-loop tools with strict schema validation and a full audit trail.
Model Context Protocol · Groq LPU · Hybrid RAG · CNN Vision
Your calendar, notes & whiteboards — one agentic brain.
A ReAct agent that searches your knowledge with dense + lexical retrieval, reads scanned images through a CNN/OCR pipeline, manages your schedule via MCP tools, and fact-checks itself before it answers.
Runs 100% in your browser. Your data and Groq key never leave your device (stored in localStorage).
12MCP tools
RRFdense + sparse fusion
<1sGroq LPU inference
2modes: browser & Docker
What it does
Hybrid RAG
Dense semantic vectors + BM25/tsvector lexical search merged by Reciprocal Rank Fusion for precise context.
CNN vision pipeline
Denoise, contrast, Sobel conv-feature maps, adaptive threshold, ROI boxes and OCR on whiteboards and scans.
Groq ReAct agent
Reason → call tools (in parallel) → observe → answer, running on Groq's ultra-fast Llama models.
Self-reflection guardrail
A second model pass checks every answer for hallucinated details and timezone/schedule mistakes.
Agent memory
Short-term, long-term and episodic memories recalled semantically across sessions.
3D context graph
Notes, events and documents as an interactive Three.js constellation linked by similarity and tags.
Production backend
TypeScript MCP server, Postgres + pgvector, Python CNN worker, all via Docker Compose.
Daily briefing & planner
AI-written briefings, conflict radar and a free-slot finder that books focus blocks for you.
Voice in English, తెలుగు, हिन्दी
Talk to your assistant and hear answers back using the browser's speech engines.
Note AI & clustering
Summaries, action items, flashcards, auto-tags and k-means topic clusters.
Multilingual agent
Ask in your language — the agent replies in kind while staying grounded in your data.
Assistant workspace
Everything below is live. Add a Groq key in Settings for the full agent; without one, chat falls back to local hybrid search.
☀ Daily briefing
Schedule, conflicts and top priorities from your data.
🕒 Smart free-slot finder
📝 Note AI
🧩 Topic clusters
Offline k-means over note embeddings.
Architecture
Tech stack
TypeScriptNode.js / ExpressMCP SDKGroq APILlama 3.3 70BPostgreSQL 16pgvector (HNSW)tsvector / pg_trgmReciprocal Rank FusionPython / FastAPIOpenCVTesseract OCRCNN feature mapsZodNext.jsTailwindThree.jsDocker Compose
About the developer
NC
Nikhil Chary Sriramoju
AI/ML & Data Engineering · Agentic AI builder
📍 Thorrur, Telangana, India
● Open to collaboration & opportunities
🎓 Education
B.Tech, Computer Science & EngineeringVaagdevi College of Engineering · 2023 – 2027 (final year)
💼 Experience
Machine Learning Engineer InternYuva Intern (NSDC) · Aug – Sep 2026
Data Analytics InternThiranex · Jul – Aug 2026
🏅 Certifications
NASSCOM · Skill India Digital Hub — AI-Data Engineering & Cloud Infrastructure AnalystIBM SkillsBuild — AI FundamentalsReliance Foundation — AI/ML Engineer Foundation
🛠 Core skills
Agentic AIRAGLLMs & LoRAComputer Vision / CNNMCPLangGraphFastAPITypeScriptNext.jsPostgreSQL / pgvectorDockerData Engineering
Selected projects
AetherNexus-Quantum
3D spatial digital twin with ViT/CNN vision, Graph-RAG and an agent swarm.
VisionMesh AI
Visual diagnostics platform: CNN vision, LangGraph agents, hybrid RAG + Neo4j.
ModelRoute Enterprise
Multimodal distillation, QLoRA and CNN-based model routing.
Go AI Gateway
Low-latency Go gateway with MCP, semantic cache and rate limiting.
LoRA Forge
LoRA fine-tuning pipeline for Llama 3 8B / Mistral 7B.
OmniLingua AI
Agentic translation app with RAG glossary, OCR and voice.
Deploy
Full stack (Docker)
cp .env.example .env # add GROQ_API_KEY
docker compose up --buildThen open Settings → set “Backend API URL” to http://localhost:4000 to route chat through the TypeScript MCP server and Postgres.
This site (GitHub Pages)
Repo → Settings → Pages
Source: Deploy from a branch
Branch: main /docsIt is pure static HTML/JS — no build step.
FAQ
Where is my data stored?
In this browser's localStorage only. Use Audit → Export data to back it up, or Delete all data to wipe it.
Is my Groq key safe?
It is kept in localStorage and sent only to api.groq.com. Anyone with access to your browser profile could read it, so use a restricted key and don't enter it on shared devices.
Why does chat work without a key?
Offline mode runs local hybrid search over your notes, documents and upcoming events. The full ReAct agent with tool calls needs a Groq key.
How does the vision pipeline work in the browser?
Canvas grayscale → contrast stretch → box-blur denoise → Sobel convolution + ReLU + pooling feature maps → adaptive threshold → Tesseract.js OCR with line regions. The Docker version uses OpenCV + Tesseract.
Are embeddings truly semantic?
The browser build uses hashed word/character n-gram vectors fused with BM25, which is strong on keywords and fuzzy matches but not deep semantics. Plug a real embedding model into the backend for that.