diff --git a/src/lib/server/ai/gemini-client.ts b/src/lib/server/ai/gemini-client.ts
new file mode 100644
index 0000000..2f45c4d
--- /dev/null
+++ b/src/lib/server/ai/gemini-client.ts
@@ -0,0 +1,150 @@
+import { GoogleGenerativeAI } from '@google/generative-ai';
+import { env } from '$env/dynamic/private';
+import {
+  RECIPE_EXTRACTION_SYSTEM_PROMPT,
+  GEMINI_RESPONSE_SCHEMA,
+  extractionResponseSchema,
+  type ExtractionResponse
+} from './recipe-extraction-prompt';
+
+export type GeminiErrorCode =
+  | 'AI_NOT_CONFIGURED'
+  | 'AI_RATE_LIMITED'
+  | 'AI_TIMEOUT'
+  | 'AI_FAILED';
+
+export class GeminiError extends Error {
+  constructor(
+    public readonly code: GeminiErrorCode,
+    message: string
+  ) {
+    super(message);
+    this.name = 'GeminiError';
+  }
+}
+
+function getStatus(err: unknown): number | undefined {
+  if (err && typeof err === 'object' && 'status' in err) {
+    const s = (err as { status?: unknown }).status;
+    if (typeof s === 'number') return s;
+  }
+  return undefined;
+}
+
+function getCfg(): { apiKey: string; model: string; timeoutMs: number } {
+  const apiKey = env.GEMINI_API_KEY ?? process.env.GEMINI_API_KEY ?? '';
+  const model =
+    env.GEMINI_MODEL ?? process.env.GEMINI_MODEL ?? 'gemini-2.5-flash';
+  const rawTimeout =
+    env.GEMINI_TIMEOUT_MS ?? process.env.GEMINI_TIMEOUT_MS ?? '20000';
+  const timeoutMs = Number(rawTimeout) || 20000;
+  return { apiKey, model, timeoutMs };
+}
+
+async function withTimeout<T>(promise: Promise<T>, ms: number): Promise<T> {
+  return new Promise((resolve, reject) => {
+    const timer = setTimeout(
+      () => reject(new GeminiError('AI_TIMEOUT', `Gemini timeout after ${ms} ms`)),
+      ms
+    );
+    promise.then(
+      (v) => {
+        clearTimeout(timer);
+        resolve(v);
+      },
+      (e) => {
+        clearTimeout(timer);
+        reject(e);
+      }
+    );
+  });
+}
+
+async function callGemini(
+  imageBuffer: Buffer,
+  mimeType: string,
+  appendUserNote?: string
+): Promise<ExtractionResponse> {
+  const { apiKey, model: modelId, timeoutMs } = getCfg();
+  if (!apiKey) {
+    throw new GeminiError('AI_NOT_CONFIGURED', 'GEMINI_API_KEY is not set');
+  }
+
+  const client = new GoogleGenerativeAI(apiKey);
+  const model = client.getGenerativeModel({
+    model: modelId,
+    systemInstruction: RECIPE_EXTRACTION_SYSTEM_PROMPT,
+    generationConfig: {
+      temperature: 0.1,
+      responseMimeType: 'application/json',
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      responseSchema: GEMINI_RESPONSE_SCHEMA as any
+    }
+  });
+
+  const parts: Array<
+    { inlineData: { data: string; mimeType: string } } | { text: string }
+  > = [{ inlineData: { data: imageBuffer.toString('base64'), mimeType } }];
+  if (appendUserNote) parts.push({ text: appendUserNote });
+
+  const result = await withTimeout(
+    model.generateContent({ contents: [{ role: 'user', parts }] }),
+    timeoutMs
+  );
+  const text = result.response.text();
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(text);
+  } catch {
+    throw new GeminiError('AI_FAILED', 'Gemini returned non-JSON output');
+  }
+  const validated = extractionResponseSchema.safeParse(parsed);
+  if (!validated.success) {
+    throw new GeminiError(
+      'AI_FAILED',
+      `Schema validation failed: ${validated.error.message}`
+    );
+  }
+  return validated.data;
+}
+
+// Public entry: one retry on recoverable failures (5xx or schema-invalid),
+// no retry on 429, AI_TIMEOUT, or config errors.
+export async function extractRecipeFromImage(
+  imageBuffer: Buffer,
+  mimeType: string
+): Promise<ExtractionResponse> {
+  try {
+    return await callGemini(imageBuffer, mimeType);
+  } catch (e) {
+    if (e instanceof GeminiError && e.code === 'AI_NOT_CONFIGURED') throw e;
+    if (e instanceof GeminiError && e.code === 'AI_TIMEOUT') throw e;
+
+    const status = getStatus(e);
+    if (status === 429) throw new GeminiError('AI_RATE_LIMITED', 'Gemini rate limit');
+
+    const recoverable =
+      (e instanceof GeminiError && e.code === 'AI_FAILED') ||
+      (status !== undefined && status >= 500);
+    if (!recoverable) {
+      throw e instanceof GeminiError
+        ? e
+        : new GeminiError('AI_FAILED', String(e));
+    }
+
+    await new Promise((r) => setTimeout(r, 500));
+    try {
+      return await callGemini(
+        imageBuffer,
+        mimeType,
+        'Dein vorheriger Output war ungültig. Bitte antworte ausschließlich mit JSON gemäß Schema.'
+      );
+    } catch (retryErr) {
+      if (retryErr instanceof GeminiError) throw retryErr;
+      const retryStatus = getStatus(retryErr);
+      if (retryStatus === 429)
+        throw new GeminiError('AI_RATE_LIMITED', 'Gemini rate limit on retry');
+      throw new GeminiError('AI_FAILED', String(retryErr));
+    }
+  }
+}
diff --git a/tests/unit/gemini-client.test.ts b/tests/unit/gemini-client.test.ts
new file mode 100644
index 0000000..93795a6
--- /dev/null
+++ b/tests/unit/gemini-client.test.ts
@@ -0,0 +1,128 @@
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+
+const mockGenerateContent = vi.fn();
+
+vi.mock('@google/generative-ai', async () => {
+  const actual = await vi.importActual<typeof import('@google/generative-ai')>(
+    '@google/generative-ai'
+  );
+  return {
+    ...actual,
+    GoogleGenerativeAI: vi.fn().mockImplementation(() => ({
+      getGenerativeModel: () => ({ generateContent: mockGenerateContent })
+    }))
+  };
+});
+
+// $env/dynamic/private is mocked via SvelteKit's own mock; here we lean on
+// process.env (the client falls back to it).
+import {
+  extractRecipeFromImage,
+  GeminiError
+} from '../../src/lib/server/ai/gemini-client';
+
+beforeEach(() => {
+  mockGenerateContent.mockReset();
+  process.env.GEMINI_API_KEY = 'test-key';
+  process.env.GEMINI_MODEL = 'gemini-2.5-flash';
+  process.env.GEMINI_TIMEOUT_MS = '5000';
+});
+
+const validResponse = {
+  title: 'Apfelkuchen',
+  servings_default: 8,
+  servings_unit: 'Stück',
+  prep_time_min: 20,
+  cook_time_min: 45,
+  total_time_min: null,
+  ingredients: [{ quantity: 500, unit: 'g', name: 'Mehl', note: null }],
+  steps: [{ text: 'Ofen auf 180 °C vorheizen.' }]
+};
+
+describe('extractRecipeFromImage', () => {
+  it('happy path: returns parsed recipe data', async () => {
+    mockGenerateContent.mockResolvedValueOnce({
+      response: { text: () => JSON.stringify(validResponse) }
+    });
+    const result = await extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg');
+    expect(result.title).toBe('Apfelkuchen');
+    expect(result.ingredients).toHaveLength(1);
+  });
+
+  it('retries once on schema-invalid JSON, then succeeds', async () => {
+    mockGenerateContent
+      .mockResolvedValueOnce({ response: { text: () => '{"title": "no arrays"}' } })
+      .mockResolvedValueOnce({
+        response: { text: () => JSON.stringify(validResponse) }
+      });
+    const result = await extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg');
+    expect(result.title).toBe('Apfelkuchen');
+    expect(mockGenerateContent).toHaveBeenCalledTimes(2);
+  });
+
+  it('throws AI_FAILED after schema-invalid + retry also invalid', async () => {
+    mockGenerateContent
+      .mockResolvedValueOnce({ response: { text: () => '{}' } })
+      .mockResolvedValueOnce({
+        response: { text: () => '{"title": "still bad"}' }
+      });
+    await expect(
+      extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg')
+    ).rejects.toMatchObject({ code: 'AI_FAILED' });
+  });
+
+  it('throws AI_RATE_LIMITED without retry on 429', async () => {
+    const err = new Error('429 Too Many Requests') as Error & { status?: number };
+    err.status = 429;
+    mockGenerateContent.mockRejectedValueOnce(err);
+    await expect(
+      extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg')
+    ).rejects.toMatchObject({ code: 'AI_RATE_LIMITED' });
+    expect(mockGenerateContent).toHaveBeenCalledTimes(1);
+  });
+
+  it('retries once on 5xx, then succeeds', async () => {
+    const err = new Error('500 Server Error') as Error & { status?: number };
+    err.status = 500;
+    mockGenerateContent
+      .mockRejectedValueOnce(err)
+      .mockResolvedValueOnce({
+        response: { text: () => JSON.stringify(validResponse) }
+      });
+    const result = await extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg');
+    expect(result.title).toBe('Apfelkuchen');
+    expect(mockGenerateContent).toHaveBeenCalledTimes(2);
+  });
+
+  it('throws AI_FAILED after 5xx + retry also fails', async () => {
+    const err = new Error('500') as Error & { status?: number };
+    err.status = 500;
+    mockGenerateContent.mockRejectedValue(err);
+    await expect(
+      extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg')
+    ).rejects.toMatchObject({ code: 'AI_FAILED' });
+    expect(mockGenerateContent).toHaveBeenCalledTimes(2);
+  });
+
+  it('throws AI_TIMEOUT when generateContent never resolves', async () => {
+    process.env.GEMINI_TIMEOUT_MS = '50';
+    mockGenerateContent.mockImplementation(
+      () => new Promise(() => {}) // never resolves
+    );
+    await expect(
+      extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg')
+    ).rejects.toMatchObject({ code: 'AI_TIMEOUT' });
+  });
+
+  it('throws AI_NOT_CONFIGURED when GEMINI_API_KEY is empty', async () => {
+    process.env.GEMINI_API_KEY = '';
+    await expect(
+      extractRecipeFromImage(Buffer.from('fake'), 'image/jpeg')
+    ).rejects.toMatchObject({ code: 'AI_NOT_CONFIGURED' });
+  });
+
+  it('GeminiError has a code property', () => {
+    const e = new GeminiError('AI_FAILED', 'x');
+    expect(e.code).toBe('AI_FAILED');
+  });
+});