Meslie - Final submission

app.py

@@ -1,60 +1,142 @@
 import gradio as gr
+from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
 import torch
-from transformers import (
-    BlipProcessor,
-    BlipForConditionalGeneration,
-    AutoTokenizer,
-    AutoModelForSeq2SeqLM
+import pypdf
+import re
+
+# NER Model: Standard setup with aggregation
+skill_pipe = pipeline(
+    "token-classification", 
+    model="ihk/skillner", 
+    aggregation_strategy="simple"
 )
 
-# Image captioning model
-caption_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
-caption_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
+# LLM: Flan-T5-Large
+model_id = "google/flan-t5-large"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
+
+def extract_text_from_pdf(file_path):
+    text = ""
+    try:
+        with open(file_path, "rb") as f:
+            reader = pypdf.PdfReader(f)
+            for page in reader.pages:
+                content = page.extract_text()
+                if content:
+                    text += content + "\n"
+    except Exception:
+        return ""
+    return text
 
-# English-to-Spanish translation model
-translation_tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-es")
-translation_model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-es")
+def analyze_profile(file_obj, text_input):
+    resume_text = ""
+    if file_obj is not None:
+        if file_obj.name.endswith('.pdf'):
+            resume_text = extract_text_from_pdf(file_obj.name)
+        else:
+            with open(file_obj.name, "r", encoding="utf-8") as f:
+                resume_text = f.read()
+    else:
+        resume_text = text_input
 
-def analyze_image(image):
-    if image is None:
-        return "Please upload an image first.", ""
+    if not resume_text.strip():
+        return "Please provide a resume.", "N/A"
 
+    # Step 1: Skill Extraction & Cleaning
     try:
-        # Generate English caption
-        inputs = caption_processor(image, return_tensors="pt")
-        caption_ids = caption_model.generate(**inputs, max_new_tokens=40)
-        english_caption = caption_processor.decode(caption_ids[0], skip_special_tokens=True)
-
-        # Translate caption to Spanish
-        translation_inputs = translation_tokenizer(
-            english_caption,
-            return_tensors="pt",
-            padding=True
-        )
-        translation_ids = translation_model.generate(**translation_inputs, max_new_tokens=60)
-        spanish_caption = translation_tokenizer.decode(
-            translation_ids[0],
-            skip_special_tokens=True
-        )
-
-        return english_caption, spanish_caption
-
-    except Exception as e:
-        return f"Error: {str(e)}", ""
-
-demo = gr.Interface(
-    fn=analyze_image,
-    inputs=gr.Image(type="pil", label="Upload an Image"),
-    outputs=[
-        gr.Textbox(label="English Image Caption"),
-        gr.Textbox(label="Spanish Translation")
-    ],
-    title="Multimodal Image Captioning and Spanish Translation Tool",
-    description=(
-        "This app uses two AI models. First, it uses a vision-language model to describe "
-        "an uploaded image in English. Then, it uses a language translation model to translate "
-        "that caption into Spanish."
+        results = skill_pipe(resume_text[:2000]) 
+        raw_skills = [res['word'].strip() for res in results if 'SKILL' in res['entity_group']]
+        
+        clean_skills = []
+        for s in raw_skills:
+            s = s.replace("##", "")
+            # Removing non-alphabetic noise
+            s = re.sub(r'^[^a-zA-Z]+|[^a-zA-Z]+$', '', s)
+            if len(s) > 2:
+                clean_skills.append(s)
+        
+        skills = sorted(list(set(clean_skills)))
+    except Exception:
+        return "Extraction failed.", "N/A"
+
+    if not skills:
+        return "No skills detected.", "Try a more detailed resume."
+
+    skills_str = ", ".join(skills)
+
+    # Step 2: Few-Shot Prompting
+
+    few_shot_prompt = (
+        "Example:\n"
+        "Skills: Python, SQL, AWS, Spark\n"
+        "Certifications:\n"
+        "1. AWS Certified Solutions Architect - Professional\n"
+        "2. Databricks Certified Data Engineer Professional\n"
+        "3. Google Professional Data Engineer\n\n"
+        f"Actual Task:\n"
+        f"Skills: {skills_str}\n"
+        "Certifications:\n"
+        "1."
     )
-)
+    
+    inputs = tokenizer(few_shot_prompt, return_tensors="pt", truncation=True, max_length=512)
+    
+
+    outputs = model.generate(
+        **inputs, 
+        max_new_tokens=80, 
+        num_beams=4,                 # Beam search is logical/less random
+        repetition_penalty=1.2,      # Lower penalty avoids hallucinations
+        no_repeat_ngram_size=2,      # Prevents word-looping
+        early_stopping=True
+    )
+    
+    roadmap = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    
+    # Cleaning the output to ensure it starts with our '1.'
+    if not roadmap.startswith("1."):
+        roadmap = "1. " + roadmap
+
+    return skills_str, roadmap
+
+# Documentation 
+arch_doc = """
+### System Architecture
+This system utilizes a three-stage pipeline for document intelligence. I integrated pypdf for raw document ingestion, followed by a JobBERT model fine-tuned for professional skill extraction. For the final reasoning layer, I used Flan-T5-Large to perform a gap analysis and generate an upskilling roadmap.
+"""
+
+eval_doc = """
+### Challenges & Solutions
+Fragmentation: To fix sub-word tokenization (e.g., "Py" + "##spark"), I implemented an Aggregation Strategy and custom Regex cleaning.
+Hallucinations: Early testing showed model "parrotting." I solved this using Few-Shot In-Context Learning and Beam Search (num_beams = 4‬) to ensure factual certification names.
+"""
+
+ethics_doc = """
+### Responsible AI & Ethics
+Bias Mitigation: The model is "feature-restricted"—the LLM only sees technical skills, blinding it to demographic data like names or gender.
+Data Privacy: This application uses transient processing; all document data exists only in volatile memory and is discarded immediately after the session ends.
+"""
+
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# Advanced Resume Intelligence")
+    
+    with gr.Row():
+        with gr.Column():
+            file_in = gr.File(label="Upload Resume (PDF/TXT)", file_types=[".pdf", ".txt"])
+            txt_in = gr.Textbox(label="Or Paste Text", lines=5)
+            btn = gr.Button("Analyze Profile", variant="primary")
+        
+        with gr.Column():
+            skill_out = gr.Textbox(label="Identified Skills")
+            advice_out = gr.Textbox(label="AI Training Roadmap")
+            
+    btn.click(fn=analyze_profile, inputs=[file_in, txt_in], outputs=[skill_out, advice_out])
+
+    with gr.Accordion("Project Documentation", open=False):
+        gr.Markdown(arch_doc)
+        gr.Markdown(eval_doc)
+        gr.Markdown(ethics_doc)
 
-demo.launch()
+if __name__ == "__main__":
+    demo.launch()