diff --git a/.gitignore b/.gitignore
index 620150d..b6930c0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -216,3 +216,5 @@ __marimo__/
 
 # Streamlit
 .streamlit/secrets.toml
+ai_engine/gateway/.env
+ai_engine/ollama/models/
diff --git a/ai_engine/README.md b/ai_engine/README.md
new file mode 100644
index 0000000..b322a4e
--- /dev/null
+++ b/ai_engine/README.md
@@ -0,0 +1,26 @@
+# LoopHarvest AI Engine
+
+Base local AI API stack for LoopHarvest using FastAPI, Ollama, and Docker Compose.
+
+## Structure
+
+```text
+ai_engine/
+  docker-compose.yml
+  gateway/
+    Dockerfile
+    requirements.txt
+    pyproject.toml
+    .env.example
+    src/
+    tests/
+  ollama/
+    models/
+```
+
+## Quick start
+
+1. Copy `gateway/.env.example` to `gateway/.env` and set `MY_API_KEY`.
+2. Run `docker compose up -d` from `ai_engine/`.
+3. Pull a model with `docker exec loopharvest-ollama ollama pull qwen2.5:1.5b`.
+4. Call `GET /health`, `GET /models`, or `POST /v1/chat` on `http://localhost:8000`.
diff --git a/ai_engine/docker-compose.yml b/ai_engine/docker-compose.yml
new file mode 100644
index 0000000..b8785d8
--- /dev/null
+++ b/ai_engine/docker-compose.yml
@@ -0,0 +1,20 @@
+services:
+  ollama:
+    image: ollama/ollama:latest
+    container_name: loopharvest-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ./ollama/models:/root/.ollama
+    restart: unless-stopped
+
+  gateway:
+    build: ./gateway
+    container_name: loopharvest-ai-gateway
+    ports:
+      - "8000:8000"
+    env_file:
+      - ./gateway/.env
+    depends_on:
+      - ollama
+    restart: unless-stopped
diff --git a/ai_engine/gateway/.env.example b/ai_engine/gateway/.env.example
new file mode 100644
index 0000000..a7a764e
--- /dev/null
+++ b/ai_engine/gateway/.env.example
@@ -0,0 +1,5 @@
+MY_API_KEY=replace-with-a-long-random-secret
+DEFAULT_MODEL=qwen2.5:1.5b
+FALLBACK_MODEL=gemma3:1b
+OLLAMA_URL=http://ollama:11434
+REQUEST_TIMEOUT_SECONDS=120
diff --git a/ai_engine/gateway/Dockerfile b/ai_engine/gateway/Dockerfile
new file mode 100644
index 0000000..08d3878
--- /dev/null
+++ b/ai_engine/gateway/Dockerfile
@@ -0,0 +1,10 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY src ./src
+
+CMD ["uvicorn", "src.main:app", "--host", "0.0.0.0", "--port", "8000"]
diff --git a/ai_engine/gateway/pyproject.toml b/ai_engine/gateway/pyproject.toml
new file mode 100644
index 0000000..5b5860e
--- /dev/null
+++ b/ai_engine/gateway/pyproject.toml
@@ -0,0 +1,34 @@
+[project]
+name = "loopharvest-ai-gateway"
+version = "0.1.0"
+description = "Local AI gateway for Ollama-backed chat requests"
+requires-python = ">=3.11"
+dependencies = [
+  "fastapi>=0.115,<1.0",
+  "httpx>=0.27,<1.0",
+  "pydantic-settings>=2.4,<3.0",
+  "python-dotenv>=1.0,<2.0",
+  "uvicorn[standard]>=0.32,<1.0",
+]
+
+[project.optional-dependencies]
+dev = [
+  "pytest>=8.3,<9.0",
+  "pytest-asyncio>=0.24,<1.0",
+]
+
+[build-system]
+requires = ["setuptools>=68", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools]
+package-dir = {"" = "src"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+pythonpath = ["."]
+testpaths = ["tests"]
+asyncio_mode = "auto"
+asyncio_default_fixture_loop_scope = "function"
diff --git a/ai_engine/gateway/requirements.txt b/ai_engine/gateway/requirements.txt
new file mode 100644
index 0000000..ab1249e
--- /dev/null
+++ b/ai_engine/gateway/requirements.txt
@@ -0,0 +1,5 @@
+fastapi>=0.115,<1.0
+httpx>=0.27,<1.0
+pydantic-settings>=2.4,<3.0
+python-dotenv>=1.0,<2.0
+uvicorn[standard]>=0.32,<1.0
diff --git a/ai_engine/gateway/src/__init__.py b/ai_engine/gateway/src/__init__.py
new file mode 100644
index 0000000..d68df01
--- /dev/null
+++ b/ai_engine/gateway/src/__init__.py
@@ -0,0 +1 @@
+"""LoopHarvest AI gateway package."""
diff --git a/ai_engine/gateway/src/config.py b/ai_engine/gateway/src/config.py
new file mode 100644
index 0000000..ab2ec7a
--- /dev/null
+++ b/ai_engine/gateway/src/config.py
@@ -0,0 +1,24 @@
+from functools import lru_cache
+
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(
+        env_file=".env",
+        env_file_encoding="utf-8",
+        extra="ignore",
+    )
+
+    app_name: str = "LoopHarvest AI Gateway"
+    app_version: str = "1.0.0"
+    ollama_url: str = Field(default="http://ollama:11434", alias="OLLAMA_URL")
+    api_key: str = Field(alias="MY_API_KEY")
+    default_model: str = Field(default="qwen2.5:1.5b", alias="DEFAULT_MODEL")
+    request_timeout_seconds: float = Field(default=120.0, alias="REQUEST_TIMEOUT_SECONDS")
+
+
+@lru_cache
+def get_settings() -> Settings:
+    return Settings()
diff --git a/ai_engine/gateway/src/dependencies.py b/ai_engine/gateway/src/dependencies.py
new file mode 100644
index 0000000..4f11b82
--- /dev/null
+++ b/ai_engine/gateway/src/dependencies.py
@@ -0,0 +1,12 @@
+from fastapi import HTTPException, status
+
+from src.config import Settings
+from src.models import ErrorDetail
+
+
+def verify_api_key(x_api_key: str | None, settings: Settings) -> None:
+    if x_api_key != settings.api_key:
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail=ErrorDetail(code="unauthorized", message="Invalid API key.").model_dump(),
+        )
diff --git a/ai_engine/gateway/src/main.py b/ai_engine/gateway/src/main.py
new file mode 100644
index 0000000..94a656f
--- /dev/null
+++ b/ai_engine/gateway/src/main.py
@@ -0,0 +1,144 @@
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
+
+import httpx
+from fastapi import Depends, FastAPI, HTTPException, Header, Request, status
+from fastapi.responses import JSONResponse
+
+from src.config import Settings, get_settings
+from src.dependencies import verify_api_key
+from src.models import ChatRequest, ErrorEnvelope, ErrorDetail, SuccessEnvelope
+
+
+def create_app(
+    settings: Settings,
+    ollama_client: httpx.AsyncClient | None = None,
+) -> FastAPI:
+    @asynccontextmanager
+    async def lifespan(_: FastAPI) -> AsyncIterator[None]:
+        managed_client: httpx.AsyncClient | None = None
+        if ollama_client is None:
+            managed_client = httpx.AsyncClient(
+                base_url=settings.ollama_url,
+                timeout=settings.request_timeout_seconds,
+            )
+            app.state.ollama_client = managed_client
+        else:
+            app.state.ollama_client = ollama_client
+
+        try:
+            yield
+        finally:
+            if managed_client is not None:
+                await managed_client.aclose()
+
+    app = FastAPI(title=settings.app_name, version=settings.app_version, lifespan=lifespan)
+    app.state.settings = settings
+    if ollama_client is not None:
+        app.state.ollama_client = ollama_client
+
+    def settings_dependency() -> Settings:
+        return app.state.settings
+
+    def require_api_key(
+        x_api_key: str | None = Header(default=None),
+        settings: Settings = Depends(settings_dependency),
+    ) -> None:
+        verify_api_key(x_api_key, settings)
+
+    async def get_ollama_client() -> httpx.AsyncClient:
+        return app.state.ollama_client
+
+    @app.exception_handler(HTTPException)
+    async def http_exception_handler(_: Request, exc: HTTPException) -> JSONResponse:
+        detail = exc.detail
+        if isinstance(detail, dict) and {"code", "message"} <= set(detail):
+            payload = ErrorEnvelope(error=ErrorDetail(**detail))
+        else:
+            payload = ErrorEnvelope(
+                error=ErrorDetail(code="http_error", message=str(detail)),
+            )
+        return JSONResponse(status_code=exc.status_code, content=payload.model_dump())
+
+    @app.exception_handler(Exception)
+    async def unhandled_exception_handler(_: Request, __: Exception) -> JSONResponse:
+        payload = ErrorEnvelope(
+            error=ErrorDetail(
+                code="internal_server_error",
+                message="An unexpected error occurred.",
+            )
+        )
+        return JSONResponse(status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, content=payload.model_dump())
+
+    @app.get("/health", response_model=SuccessEnvelope)
+    async def health() -> SuccessEnvelope:
+        return SuccessEnvelope(
+            data={
+                "status": "ok",
+                "service": settings.app_name,
+                "version": settings.app_version,
+            }
+        )
+
+    @app.get("/models", response_model=SuccessEnvelope)
+    async def list_models(
+        _: None = Depends(require_api_key),
+        client: httpx.AsyncClient = Depends(get_ollama_client),
+    ) -> SuccessEnvelope:
+        data = await _ollama_get(client, "/api/tags")
+        return SuccessEnvelope(data=data)
+
+    @app.post("/v1/chat", response_model=SuccessEnvelope)
+    async def chat(
+        body: ChatRequest,
+        _: None = Depends(require_api_key),
+        client: httpx.AsyncClient = Depends(get_ollama_client),
+    ) -> SuccessEnvelope:
+        payload = {
+            "model": body.model or settings.default_model,
+            "messages": [message.model_dump() for message in body.messages],
+            "stream": False,
+            "options": {
+                "temperature": body.temperature,
+                "num_predict": body.max_tokens,
+            },
+        }
+        if body.system is not None:
+            payload["system"] = body.system
+
+        data = await _ollama_post(client, "/api/chat", payload)
+        return SuccessEnvelope(
+            data={
+                "model": data.get("model", payload["model"]),
+                "content": data.get("message", {}).get("content", ""),
+            }
+        )
+
+    return app
+
+
+async def _ollama_get(client: httpx.AsyncClient, path: str) -> dict:
+    try:
+        response = await client.get(path)
+        response.raise_for_status()
+        return response.json()
+    except (httpx.HTTPError, ValueError) as exc:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=ErrorDetail(code="ollama_error", message="Ollama request failed.").model_dump(),
+        ) from exc
+
+
+async def _ollama_post(client: httpx.AsyncClient, path: str, payload: dict) -> dict:
+    try:
+        response = await client.post(path, json=payload)
+        response.raise_for_status()
+        return response.json()
+    except (httpx.HTTPError, ValueError) as exc:
+        raise HTTPException(
+            status_code=status.HTTP_502_BAD_GATEWAY,
+            detail=ErrorDetail(code="ollama_error", message="Ollama request failed.").model_dump(),
+        ) from exc
+
+
+app = create_app(get_settings())
diff --git a/ai_engine/gateway/src/models.py b/ai_engine/gateway/src/models.py
new file mode 100644
index 0000000..0d0dbc0
--- /dev/null
+++ b/ai_engine/gateway/src/models.py
@@ -0,0 +1,31 @@
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+class ErrorDetail(BaseModel):
+    code: str
+    message: str
+
+
+class SuccessEnvelope(BaseModel):
+    success: Literal[True] = True
+    data: dict[str, Any]
+
+
+class ErrorEnvelope(BaseModel):
+    success: Literal[False] = False
+    error: ErrorDetail
+
+
+class ChatMessage(BaseModel):
+    role: str = Field(min_length=1, max_length=32)
+    content: str = Field(min_length=1)
+
+
+class ChatRequest(BaseModel):
+    messages: list[ChatMessage] = Field(min_length=1)
+    model: str | None = Field(default=None, min_length=1)
+    max_tokens: int = Field(default=1024, ge=1, le=16384)
+    temperature: float = Field(default=0.7, ge=0.0, le=2.0)
+    system: str | None = Field(default=None, min_length=1)
diff --git a/ai_engine/gateway/src/utils/image.py b/ai_engine/gateway/src/utils/image.py
new file mode 100644
index 0000000..d3560c9
--- /dev/null
+++ b/ai_engine/gateway/src/utils/image.py
@@ -0,0 +1,24 @@
+import base64, httpx
+from pathlib import Path
+
+async def to_base64(source: str) -> str:
+    """Accept a URL, file path, or raw base64 string — always return base64."""
+    
+    # Already base64
+    if source.startswith("data:image"):
+        return source.split(",")[1]
+    
+    # Remote URL — download it
+    if source.startswith("http"):
+        async with httpx.AsyncClient(timeout=30) as client:
+            r = await client.get(source)
+            return base64.b64encode(r.content).decode("utf-8")
+    
+    # Local file path
+    return base64.b64encode(Path(source).read_bytes()).decode("utf-8")
+
+def detect_media_type(source: str) -> str:
+    if ".png" in source:  return "image/png"
+    if ".gif" in source:  return "image/gif"
+    if ".webp" in source: return "image/webp"
+    return "image/jpeg"   # safe default
\ No newline at end of file
diff --git a/ai_engine/gateway/tests/conftest.py b/ai_engine/gateway/tests/conftest.py
new file mode 100644
index 0000000..d0f99eb
--- /dev/null
+++ b/ai_engine/gateway/tests/conftest.py
@@ -0,0 +1,28 @@
+import os
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+
+os.environ.setdefault("MY_API_KEY", "test-secret")
+
+from src.config import Settings
+from src.main import create_app
+
+
+@pytest.fixture
+def settings() -> Settings:
+    return Settings(
+        app_name="LoopHarvest AI Gateway",
+        api_key="test-secret",
+        default_model="qwen2.5:1.5b",
+        ollama_url="http://ollama:11434",
+        request_timeout_seconds=30.0,
+    )
+
+
+@pytest.fixture
+async def client(settings: Settings) -> AsyncClient:
+    app = create_app(settings)
+    transport = ASGITransport(app=app)
+    async with AsyncClient(transport=transport, base_url="http://testserver") as async_client:
+        yield async_client
diff --git a/ai_engine/gateway/tests/test_gateway.py b/ai_engine/gateway/tests/test_gateway.py
new file mode 100644
index 0000000..8d47b68
--- /dev/null
+++ b/ai_engine/gateway/tests/test_gateway.py
@@ -0,0 +1,131 @@
+import pytest
+from fastapi import status
+from httpx import ASGITransport, AsyncClient
+
+from src.config import Settings
+from src.main import create_app
+
+
+class OllamaMockTransport(ASGITransport):
+    def __init__(self) -> None:
+        from fastapi import FastAPI, Response
+
+        app = FastAPI()
+
+        @app.get("/api/tags")
+        async def list_models():
+            return {"models": [{"name": "qwen2.5:1.5b"}, {"name": "gemma3:1b"}]}
+
+        @app.post("/api/chat")
+        async def chat(payload: dict):
+            if payload["model"] == "missing-model":
+                return Response(
+                    content='{"error":"model not found"}',
+                    media_type="application/json",
+                    status_code=status.HTTP_404_NOT_FOUND,
+                )
+
+            return {
+                "model": payload["model"],
+                "message": {"content": "mocked reply"},
+            }
+
+        super().__init__(app=app)
+
+
+@pytest.fixture
+async def upstream_client() -> AsyncClient:
+    async with AsyncClient(transport=OllamaMockTransport(), base_url="http://ollama") as client:
+        yield client
+
+
+@pytest.fixture
+async def app_client(settings: Settings, upstream_client: AsyncClient) -> AsyncClient:
+    app = create_app(settings, ollama_client=upstream_client)
+    transport = ASGITransport(app=app)
+    async with AsyncClient(transport=transport, base_url="http://testserver") as client:
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_health_endpoint_is_public(client: AsyncClient) -> None:
+    response = await client.get("/health")
+
+    assert response.status_code == status.HTTP_200_OK
+    assert response.json() == {
+        "success": True,
+        "data": {
+            "status": "ok",
+            "service": "LoopHarvest AI Gateway",
+            "version": "1.0.0",
+        },
+    }
+
+
+@pytest.mark.asyncio
+async def test_models_requires_api_key(client: AsyncClient) -> None:
+    response = await client.get("/models")
+
+    assert response.status_code == status.HTTP_401_UNAUTHORIZED
+    assert response.json() == {
+        "success": False,
+        "error": {
+            "code": "unauthorized",
+            "message": "Invalid API key.",
+        },
+    }
+
+
+@pytest.mark.asyncio
+async def test_models_returns_upstream_models(app_client: AsyncClient) -> None:
+    response = await app_client.get("/models", headers={"x-api-key": "test-secret"})
+
+    assert response.status_code == status.HTTP_200_OK
+    assert response.json() == {
+        "success": True,
+        "data": {
+            "models": [{"name": "qwen2.5:1.5b"}, {"name": "gemma3:1b"}],
+        },
+    }
+
+
+@pytest.mark.asyncio
+async def test_chat_uses_default_model_and_returns_content(app_client: AsyncClient) -> None:
+    response = await app_client.post(
+        "/v1/chat",
+        headers={"x-api-key": "test-secret"},
+        json={
+            "messages": [{"role": "user", "content": "Hello"}],
+            "temperature": 0.5,
+        },
+    )
+
+    assert response.status_code == status.HTTP_200_OK
+    assert response.json() == {
+        "success": True,
+        "data": {
+            "model": "qwen2.5:1.5b",
+            "content": "mocked reply",
+        },
+    }
+
+
+@pytest.mark.asyncio
+async def test_chat_surfaces_upstream_not_found_as_bad_gateway(app_client: AsyncClient) -> None:
+    response = await app_client.post(
+        "/v1/chat",
+        headers={"x-api-key": "test-secret"},
+        json={
+            "model": "missing-model",
+            "messages": [{"role": "user", "content": "Hello"}],
+        },
+    )
+
+    assert response.status_code == status.HTTP_502_BAD_GATEWAY
+    assert response.json() == {
+        "success": False,
+        "error": {
+            "code": "ollama_error",
+            "message": "Ollama request failed.",
+        },
+    }