diff --git a/internal/aiprocess/aiprocess.go b/internal/aiprocess/aiprocess.go
new file mode 100644
index 0000000..7725d61
--- /dev/null
+++ b/internal/aiprocess/aiprocess.go
@@ -0,0 +1,247 @@
+package aiprocess
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+
+	_ "embed"
+)
+
+type Config struct {
+	Enabled          bool
+	Provider         string
+	Model            string
+	Temperature      float64
+	SystemPromptFile string
+	BaseURL          string
+	APIKey           string
+	TimeoutSec       int
+}
+
+type Processor interface {
+	Process(ctx context.Context, input string) (string, error)
+}
+
+func New(cfg Config) (Processor, error) {
+	if !cfg.Enabled {
+		return nil, nil
+	}
+
+	provider := strings.ToLower(strings.TrimSpace(cfg.Provider))
+	if provider == "" {
+		return nil, errors.New("ai provider is required when enabled")
+	}
+	if strings.TrimSpace(cfg.Model) == "" {
+		return nil, errors.New("ai model is required when enabled")
+	}
+
+	systemPrompt, err := loadSystemPrompt(cfg.SystemPromptFile)
+	if err != nil {
+		return nil, err
+	}
+
+	timeout := time.Duration(cfg.TimeoutSec) * time.Second
+	if timeout <= 0 {
+		timeout = 20 * time.Second
+	}
+
+	switch provider {
+	case "ollama":
+		base := strings.TrimRight(cfg.BaseURL, "/")
+		if base == "" {
+			base = "http://localhost:11434"
+		}
+		return &ollamaProcessor{
+			client:      &http.Client{Timeout: timeout},
+			baseURL:     base,
+			model:       cfg.Model,
+			temperature: cfg.Temperature,
+			system:      systemPrompt,
+		}, nil
+	case "openai_compat":
+		base := strings.TrimRight(cfg.BaseURL, "/")
+		if base == "" {
+			return nil, errors.New("ai base_url is required for openai_compat")
+		}
+		return &openAICompatProcessor{
+			client:      &http.Client{Timeout: timeout},
+			baseURL:     base,
+			apiKey:      cfg.APIKey,
+			model:       cfg.Model,
+			temperature: cfg.Temperature,
+			system:      systemPrompt,
+		}, nil
+	default:
+		return nil, fmt.Errorf("unknown ai provider %q", provider)
+	}
+}
+
+func loadSystemPrompt(path string) (string, error) {
+	if strings.TrimSpace(path) == "" {
+		return strings.TrimSpace(defaultSystemPrompt), nil
+	}
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return "", fmt.Errorf("read system prompt file: %w", err)
+	}
+	return strings.TrimSpace(string(data)), nil
+}
+
+//go:embed system_prompt.txt
+var defaultSystemPrompt string
+
+type ollamaProcessor struct {
+	client      *http.Client
+	baseURL     string
+	model       string
+	temperature float64
+	system      string
+}
+
+func (p *ollamaProcessor) Process(ctx context.Context, input string) (string, error) {
+	reqBody := ollamaRequest{
+		Model:  p.model,
+		Prompt: input,
+		Stream: false,
+	}
+	if p.system != "" {
+		reqBody.System = p.system
+	}
+	if p.temperature != 0 {
+		reqBody.Options = &ollamaOptions{Temperature: p.temperature}
+	}
+
+	payload, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, p.baseURL+"/api/generate", bytes.NewReader(payload))
+	if err != nil {
+		return "", err
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := p.client.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("ollama request failed: %s", readErrorBody(resp.Body))
+	}
+
+	var out ollamaResponse
+	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(out.Response), nil
+}
+
+type ollamaRequest struct {
+	Model   string         `json:"model"`
+	Prompt  string         `json:"prompt"`
+	System  string         `json:"system,omitempty"`
+	Stream  bool           `json:"stream"`
+	Options *ollamaOptions `json:"options,omitempty"`
+}
+
+type ollamaOptions struct {
+	Temperature float64 `json:"temperature,omitempty"`
+}
+
+type ollamaResponse struct {
+	Response string `json:"response"`
+}
+
+type openAICompatProcessor struct {
+	client      *http.Client
+	baseURL     string
+	apiKey      string
+	model       string
+	temperature float64
+	system      string
+}
+
+func (p *openAICompatProcessor) Process(ctx context.Context, input string) (string, error) {
+	messages := []openAIMessage{
+		{Role: "user", Content: input},
+	}
+	if p.system != "" {
+		messages = append([]openAIMessage{{Role: "system", Content: p.system}}, messages...)
+	}
+
+	reqBody := openAIRequest{
+		Model:       p.model,
+		Messages:    messages,
+		Temperature: p.temperature,
+	}
+
+	payload, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", err
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, p.baseURL+"/v1/chat/completions", bytes.NewReader(payload))
+	if err != nil {
+		return "", err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if strings.TrimSpace(p.apiKey) != "" {
+		req.Header.Set("Authorization", "Bearer "+p.apiKey)
+	}
+
+	resp, err := p.client.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("openai_compat request failed: %s", readErrorBody(resp.Body))
+	}
+
+	var out openAIResponse
+	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
+		return "", err
+	}
+	if len(out.Choices) == 0 {
+		return "", errors.New("openai_compat response missing choices")
+	}
+	return strings.TrimSpace(out.Choices[0].Message.Content), nil
+}
+
+type openAIRequest struct {
+	Model       string          `json:"model"`
+	Messages    []openAIMessage `json:"messages"`
+	Temperature float64         `json:"temperature,omitempty"`
+}
+
+type openAIMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+type openAIResponse struct {
+	Choices []openAIChoice `json:"choices"`
+}
+
+type openAIChoice struct {
+	Message openAIMessage `json:"message"`
+}
+
+func readErrorBody(r io.Reader) string {
+	data, err := io.ReadAll(io.LimitReader(r, 64*1024))
+	if err != nil {
+		return "unknown error"
+	}
+	return strings.TrimSpace(string(data))
+}
diff --git a/internal/aiprocess/system_prompt.txt b/internal/aiprocess/system_prompt.txt
new file mode 100644
index 0000000..5f3722a
--- /dev/null
+++ b/internal/aiprocess/system_prompt.txt
@@ -0,0 +1,16 @@
+You are a deterministic text transcription cleaning engine.
+You transform speech transcripts into clean written text while keeping its meaning.
+
+Follow these rules strictly:
+1. Remove filler words (um, uh, like, okay so).
+2. Resolve self-corrections by keeping ONLY the final version.
+   Examples:
+   - "schedule that for 5 PM, I mean 4 PM" -> "schedule that for 4 PM"
+   - "let's ask Bob, I mean Janice, let's ask Janice" -> "let's ask Janice"
+3. Fix grammar, capitalization, and punctuation.
+4. Do NOT add new content.
+5. Do NOT remove real content.
+6. Do NOT rewrite stylistically.
+7. Preserve meaning exactly.
+
+Return ONLY the cleaned text. No explanations.
diff --git a/internal/daemon/daemon.go b/internal/daemon/daemon.go
index e934d01..4da3421 100644
--- a/internal/daemon/daemon.go
+++ b/internal/daemon/daemon.go
@@ -218,7 +218,8 @@ func (d *Daemon) stopAndProcess(reason string) {
 	}
 	d.log.Printf("transcript: %s", text)
 
-	if d.ai != nil && d.cfg.AIEnabled {
+	if d.cfg.AIEnabled && d.ai != nil {
+		d.log.Printf("ai enabled")
 		d.setState(StateProcessing)
 		aiCtx, cancel := context.WithTimeout(context.Background(), time.Duration(d.cfg.AITimeoutSec)*time.Second)
 		cleaned, err := d.ai.Process(aiCtx, text)
diff --git a/internal/ui/icons.go b/internal/ui/icons.go
index ee24b68..aec61c0 100644
--- a/internal/ui/icons.go
+++ b/internal/ui/icons.go
@@ -22,3 +22,7 @@ func IconRecording() []byte {
 func IconTranscribing() []byte {
 	return iconTranscribing
 }
+
+func IconProcessing() []byte {
+	return iconProcessing
+}