SYS_OVERVIEW / Identity_01

ALLAN_PEREZ_FELDMAN

PRINCIPAL_ENGINEER

alternate_emailAllanPerezFeldman@gmail.comarrow_forward https://github.com/Ryuketsukamiarrow_forward LINKEDIN_PROFILEarrow_forward downloadDOWNLOAD_CV

TECHNICAL_STACK

code

_STACK/WEB

TypescriptNextJSTailwindThree.js

psychology

_STACK/AGENTIC

LangChainVectorDBPythonLLM_Orch

dns

_STACK/CLOUD

DockerAWS_S3Rust_Lambda

settings_input_component

INFRA

TerraformAnsible

PROJECT_INDEX

lan

AGENT_INFRA

LangGraphONNX

memory

RUST_KERNEL

PYTS

settings_ethernet

WASM_ENGINE

layers

NUXT_CORE

ALLAN_PEREZ_FELDMAN

PRINCIPAL_ENGINEER

LOC: 52.5200° N—v7.2.0_STABLE

Technical_Stack_Architecture

_STACK/WEB

TypescriptCORE

NextJSFRAMEWORK

TailwindSTYLING

Three.js3D

_STACK/AGENTIC

LangChainLLM_OPS

VectorDBMEMORY

PythonTENSOR

LLM_OrchORCHESTRATION

_STACK/CLOUD

DockerISOLATION

AWS_S3STORAGE

Rust_LambdaCOMPUTE

INFRA

TerraformHCL

AnsiblePLAYBOOK

DEPL_2025.04.05

Agent Infrastructure Comparison

Local Machine vs. GPU Renting vs. API vs. Client Browser

Context_

Side-by-side comparison of a LangGraph ReAct agent running identical model weights across four infrastructure types: local desktop GPU via Ollama, rented cloud GPU on Vast.ai, API proxy through Railway + OpenRouter, and browser-native inference via Transformers.js.

Architecture

ReAct + SSE

Latency

varies

Status

LIVE

Deployments

Qwen3 8B

Shared Model

0 deps

Browser Runtime

The Problem Space_

Evaluating LLM deployment infrastructure is difficult because configurations differ across too many dimensions simultaneously. This project controls for model and tooling, varying only the compute substrate.

Four deployment types run the same ReAct loop against identical queries. Latency, time-to-first-token, throughput, and cost are captured per-run and displayed in real time.

LIVE_DEMO

// identical ReAct agent · four infrastructure types

LOCAL_DESKTOP

OFFLINE

AGENT_OFFLINE

Requires active desktop tunnel

langgraph dev --tunnel

TTFT

—

TOTAL

—

T/S

—

VAST_AI_GPU

OFFLINE

AGENT_OFFLINE

Scheduled: Sun–Thu 06:30–17:00

GitHub Actions managed

TTFT

—

TOTAL

—

T/S

—

OPENROUTER_API

OFFLINE

AGENT_OFFLINE

Railway deployment inactive

Always-on when deployed

TTFT

—

TOTAL

—

T/S

—

BROWSER_CLIENTQwen2.5-0.5B · ReAct

IDLE

// ReAct agent · web_search tool // type a query and press execute

TTFT

—

TOTAL

—

T/S

—

// browser column: Qwen2.5-0.5B · Transformers.js · WebGPU → WASM · web_search via /api/agent/search · model cached after first load