diff --git a/src/services/llm/llm.service.ts b/src/services/llm/llm.service.ts
new file mode 100644
index 0000000..debafa6
--- /dev/null
+++ b/src/services/llm/llm.service.ts
@@ -0,0 +1,168 @@
+import { z } from 'zod';
+import { env } from '../../config/env.js';
+import type { Stack, Level, QuestionType } from '../../db/schema/enums.js';
+
+export interface LlmConfig {
+  baseUrl: string;
+  model: string;
+  apiKey?: string;
+  timeoutMs: number;
+  temperature: number;
+  maxTokens: number;
+}
+
+export interface ChatMessage {
+  role: 'system' | 'user' | 'assistant';
+  content: string;
+}
+
+export interface ChatCompletionResponse {
+  choices: Array<{
+    message?: { content: string };
+    text?: string;
+  }>;
+}
+
+const QUESTION_TYPES: QuestionType[] = ['single_choice', 'multiple_select', 'true_false', 'short_text'];
+
+const optionSchema = z.object({
+  key: z.string().min(1),
+  text: z.string().min(1),
+});
+
+const generatedQuestionSchema = z.object({
+  questionText: z.string().min(1),
+  type: z.enum(QUESTION_TYPES as [string, ...string[]]),
+  options: z.array(optionSchema).optional(),
+  correctAnswer: z.union([z.string(), z.array(z.string())]),
+  explanation: z.string().min(1),
+});
+
+const generateQuestionsResponseSchema = z.object({
+  questions: z.array(generatedQuestionSchema),
+});
+
+export type GeneratedQuestion = z.infer<typeof generatedQuestionSchema> & {
+  stack: Stack;
+  level: Level;
+};
+
+export interface GenerateQuestionsInput {
+  stack: Stack;
+  level: Level;
+  count: number;
+  types?: QuestionType[];
+}
+
+export class LlmService {
+  private readonly config: LlmConfig;
+
+  constructor(config?: Partial<LlmConfig>) {
+    this.config = {
+      baseUrl: config?.baseUrl ?? env.LLM_BASE_URL,
+      model: config?.model ?? env.LLM_MODEL,
+      apiKey: config?.apiKey ?? env.LLM_API_KEY,
+      timeoutMs: config?.timeoutMs ?? env.LLM_TIMEOUT_MS,
+      temperature: config?.temperature ?? env.LLM_TEMPERATURE,
+      maxTokens: config?.maxTokens ?? env.LLM_MAX_TOKENS,
+    };
+  }
+
+  async chat(messages: ChatMessage[]): Promise<string> {
+    const url = `${this.config.baseUrl.replace(/\/$/, '')}/chat/completions`;
+
+    const headers: Record<string, string> = {
+      'Content-Type': 'application/json',
+    };
+    if (this.config.apiKey) {
+      headers['Authorization'] = `Bearer ${this.config.apiKey}`;
+    }
+
+    const body = {
+      model: this.config.model,
+      messages: messages.map((m) => ({ role: m.role, content: m.content })),
+      temperature: this.config.temperature,
+      max_tokens: this.config.maxTokens,
+    };
+
+    const controller = new AbortController();
+    const timeoutId = setTimeout(() => controller.abort(), this.config.timeoutMs);
+
+    try {
+      const res = await fetch(url, {
+        method: 'POST',
+        headers,
+        body: JSON.stringify(body),
+        signal: controller.signal,
+      });
+
+      clearTimeout(timeoutId);
+
+      if (!res.ok) {
+        const text = await res.text();
+        throw new Error(`LLM request failed: ${res.status} ${res.statusText} - ${text}`);
+      }
+
+      const data = (await res.json()) as ChatCompletionResponse;
+
+      const choice = data.choices?.[0];
+      const content = choice?.message?.content ?? choice?.text ?? '';
+
+      return content.trim();
+    } catch (err) {
+      clearTimeout(timeoutId);
+      if (err instanceof Error) {
+        throw err;
+      }
+      throw new Error('LLM request failed');
+    }
+  }
+
+  async generateQuestions(input: GenerateQuestionsInput): Promise<GeneratedQuestion[]> {
+    const { stack, level, count, types = QUESTION_TYPES } = input;
+
+    const typeList = types.join(', ');
+    const systemPrompt = `You are a technical interview question generator. Generate exactly ${count} programming/tech questions.
+Return ONLY valid JSON in this exact format (no markdown, no code blocks):
+{"questions":[{"questionText":"...","type":"single_choice|multiple_select|true_false|short_text","options":[{"key":"a","text":"..."}],"correctAnswer":"a" or ["a","b"],"explanation":"..."}]}
+Rules: type must be one of: ${typeList}. For single_choice/multiple_select: options array required with key (a,b,c,d). For true_false: options [{"key":"true","text":"True"},{"key":"false","text":"False"}]. For short_text: options omitted, correctAnswer is string.`;
+
+    const userPrompt = `Generate ${count} questions for stack="${stack}", level="${level}". Use types: ${typeList}.`;
+
+    const raw = await this.chat([
+      { role: 'system', content: systemPrompt },
+      { role: 'user', content: userPrompt },
+    ]);
+
+    const jsonStr = extractJson(raw);
+    const parsed = JSON.parse(jsonStr) as unknown;
+
+    const result = generateQuestionsResponseSchema.safeParse(parsed);
+    if (!result.success) {
+      throw new Error(`LLM response validation failed: ${result.error.message}`);
+    }
+
+    const questions: GeneratedQuestion[] = result.data.questions.map((q) => ({
+      ...q,
+      stack,
+      level,
+    }));
+
+    for (const q of questions) {
+      if ((q.type === 'single_choice' || q.type === 'multiple_select') && (!q.options || q.options.length === 0)) {
+        throw new Error(`Question validation failed: ${q.type} requires options`);
+      }
+      if (q.type === 'true_false' && (!q.options || q.options.length < 2)) {
+        throw new Error(`Question validation failed: true_false requires true/false options`);
+      }
+    }
+
+    return questions;
+  }
+}
+
+function extractJson(text: string): string {
+  const trimmed = text.trim();
+  const match = trimmed.match(/\{[\s\S]*\}/);
+  return match ? match[0]! : trimmed;
+}