Efrain Garay EfrainGaray

🔬 Research Projects

PhantomRelay

Production-grade stealth browser relay that defeats Cloudflare, DataDome, and PerimeterX at the TLS layer. Uses Chrome's own BoringSSL (via Rust native addon) to produce a JA4 fingerprint identical to real Chrome. Decomposed into 5 focused microservices with ISolver CAPTCHA strategy pattern, Redis distributed rate limiting, EventEmitter-based proxy health monitoring, and Docker remote Chrome support.

Architecture (Phase 11 — hardened):

5-service decomposition: BlockDetector · ProxyController · PageActionExecutor · CaptchaOrchestrator · ChromeSessionManager
ISolver Strategy pattern: pluggable CAPTCHA solvers (reCAPTCHA, Turnstile, DataDome, PerimeterX)
Redis distributed rate limiter with per-endpoint granularity (scrape: 10/min, session: 30/min)
EventEmitter on HealthTracker/ProxyPool (degraded/recovered/banned/exhausted events)
Interface segregation: SessionRef · ProfileRef · ProxyRef
Timing-safe auth, SSRF protection, WebSocket CDP auth

Key results:

✅ JA4: t13d1514h2_8daaf6152771_827b515c4f52 — verified live vs Chrome 124
✅ HTTP/2 SETTINGS exact Chrome values (headerTableSize=65536, windowSize=6291456)
✅ 13 JS stealth patches via single CDP IIFE — navigator.webdriver → undefined
✅ Profile aging: direct SQLite writes to Chrome History (6-month fake personas)
✅ bot.sannysoft.com stealth score: 31/31 (100%)
✅ 732 tests (25 integration, 57 service, 8 error-path) · TypeScript strict · MIT

Agatha

Binary data transport via YouTube VP9 4K. Encodes arbitrary files into grayscale video frames (1 bit/pixel), uploads to YouTube, and recovers with SHA256-verified fidelity using BCH(t=16) error correction.

Key results:

✅ 300 MB recovered — SHA256 verified
✅ 30 fps & 60 fps validated end-to-end
✅ BER = 0 stealth channel (C16 mode)
✅ Academic paper published

NEXCOMP

Adaptive lossless compressor with per-block codec selection. Outperforms bzip2 across standard benchmarks with 3× faster decompression.

Corpus	NEXCOMP	bzip2
Calgary	2.072 bpb	2.109 bpb
Canterbury	1.278 bpb	1.545 bpb
Silesia	2.022 bpb	2.057 bpb

Decompression: ~50 MB/s vs bzip2 ~17 MB/s

📦 Notable Open Source

ragcore

Pure RAG-as-a-service — no LLM, no opinions. Any AI (Claude, GPT-4, Llama, custom agents) queries ragcore to get grounded context from your documents via MCP or the OpenAI-compatible API. Pluggable embedding and reranker backends.

Key features:

MCP server (SSE) + OpenAI-compatible REST API — integrates with Claude Code, Cursor, LangChain, any SDK
Swappable embedders: local sentence-transformers or any /v1/embeddings endpoint (HuggingFace, NVIDIA NIM, OpenAI, Ollama…)
Swappable rerankers: local CrossEncoder or remote /v1/rerank (Cohere, Jina AI, Voyage AI)
LRU cache (300s TTL) · rate limiting (60 req/min) · namespaces · observability headers
75 tests, zero ML downloads required — runs in ~2.5s

python-fastapi-IA-rag-generator

FastAPI multi-model RAG server with LangChain, vector store, and multi-backend AI generation. Supports swappable LLM backends and retrieval-augmented pipelines.