Home/Projects/DAVAI: Autonomous Agentic Runtime with 4-Tier Memory & Dynamic Tool RAG
Agentic AI & SystemsAutonomous Agent Systems & Infrastructure Featured

DAVAI: Autonomous Agentic Runtime with 4-Tier Memory & Dynamic Tool RAG

Architected an autonomous, stateful desktop AI agent runtime decoupling LLM reasoning from execution via an asynchronous EventBus, dynamic multi-model routing (Groq, DeepSeek, Ollama), and zero-delay WebSocket streaming.

TTFT Reduction

70% Faster

Cache Performance

0ms Ephemeral LRU Cache Hits

Memory Footprint

<50MB Ephemeral RAM

Streaming Protocol

Zero-delay SSE & WebSockets

Code Repository

Abstract & Overview

DAVAI is an autonomous, stateful desktop AI agent runtime decoupling LLM reasoning from task execution. Features an asynchronous EventBus, dynamic multi-model routing (Groq, DeepSeek, Ollama), zero-delay WebSocket streaming, and a 4-tier hybrid memory engine (ephemeral RAM, ChromaDB vector embeddings, 2-hop Knowledge Graph triples, and WAL SQLite).

Problem Formulation

Monolithic agent architectures suffer from tool schema bloat, context window exhaustion, and excessive Time-to-First-Token (TTFT) when scaling to dozens of local OS tools.

Architecture & Key Mechanisms

14-Tier Memory Hierarchy: Ephemeral RAM working memory (45m sliding TTL) + ChromaDB vector embeddings + 2-hop Knowledge Graph triples + SQLite WAL ground truth.
2Dynamic Tool RAG 2.0: Prunes 19+ tool schemas down to candidate functions on-demand using hybrid lexical and vector retrieval.
3Deterministic Prefix Caching: 10-layer prompt sequencer stabilizing prefix bit-patterns to maximize cloud KV prompt cache reuse.
4Multimodal Research Pipelines: Headless Playwright Chromium web crawling coupled with targeted Tesseract OCR heuristics.

Empirical Highlights

  • Formulated a 4-tier hybrid memory engine fusing ephemeral RAM working memory (45m sliding TTL), ChromaDB vector embeddings, 2-hop Knowledge Graph triples, and SQLite (WAL mode) ground truth.
  • Engineered Dynamic Tool RAG 2.0 with a balanced-brace nested parser, dynamically pruning 19+ tool schemas to top candidate functions using hybrid vector and lexical retrieval.
  • Implemented 10-layer deterministic prompt sequencing and an ephemeral LRU cache (<50MB RAM), stabilizing prefix bit-patterns to maximize cloud KV prompt cache reuse.
  • Constructed multimodal deep research and vision pipelines integrating headless Playwright Chromium web crawling with targeted Tesseract OCR heuristics.

Technologies & Tooling

Python 3.13FastAPIWebSocketsDynamic Tool RAGChromaDBKnowledge GraphsSQLite (WAL)PlaywrightTesseract OCRGroq / LLaMA 3.3DeepSeekOllama