Final Project

Semester: Fall 2026-2027

Grade Value: 9 points

Project Topic Selection Guide

Students may choose one of the topics listed in this document and define their project accordingly. Students may also propose a new topic; in this case, the instructor will determine the appropriate level.

Project Leveling Based on Team Size

Project Level Team Size Description
Level 1 1 person Individual projects
Level 2 2 persons Two-person projects
Level 3 3 persons Three-person projects

Important Notes:

  • A student with a one-person team is permitted to undertake Level 2 and Level 3 projects.
  • However, multi-person teams cannot choose lower-level projects (e.g., Level 1 with two people).

Final Reminders

  • Topic selection must be announced to the instructors by October 23, 2026.
  • The final report must include: design, implementation, evaluation, complete documentation, and expected outputs.
  • Projects must be presented in the final session.
  • Multi-person teams are required to ensure that all members actively and effectively participate in all stages of the project (design, implementation, evaluation, documentation, and presentation).
  • At the end of the semester, each team member must submit a separate individual performance report specifying exactly which parts they were responsible for.
  • Each topic from the provided list (or proposed by students), regardless of level (1, 2, or 3) and regardless of team size, can be selected by a maximum of 2 teams.

Strong Recommendation: Clear division of labor and use of project management tools (such as GitHub Projects) to track tasks and progress is highly recommended to avoid any ambiguity or disputes.

Topic List

Level 1

Level 2

Level 3

Level 1 Topics

Topic 1: Autonomous Agent System with “Doubt” and “Reconsideration” Capabilities

Core Challenge:

Current systems are either overconfident (hallucination) or lack confidence entirely. You need to design a system that:

  • Detects when it has made a mistake (without needing external feedback)
  • Can reconsider its reasoning path
  • Admits mistakes and takes a new path
  • Changes its opinion if new evidence is found

Suggested Capabilities:

  1. Logical Doubt Module:
    • After each answer, an internal critic evaluates the response from 3 different angles:
      • Contradiction (with previous answers)
      • Misalignment with basic knowledge
      • Reasoning weakness (gaps in the logical chain)
    • If the Critic scores below 70%, the Agent enters reconsideration mode.
  2. Reconsideration Mode:
    • Agent completely forgets the previous answer (Reset Context)
    • Uses a completely different method (e.g., if previously CoT, now uses ToT)
    • Compares the new answer with the previous one and selects the best
  3. Error Memory:
    • Stores all mistakes with details:
      • What was the question?
      • What mistake was made?
      • What was the cause? (insufficient data, faulty reasoning, bias)
    • When facing similar questions, warns: “I’ve made mistakes on this topic before, let’s be more careful”
  4. Dynamic Confidence Metric:
    • Each answer has a confidence score (0 to 100)
    • If score is below 50, start response with phrases like “I’m not sure, but…”
    • If score is above 90, respond with full confidence
  5. Devil’s Advocate Simulation:
    • Create an adversary agent that always tries to refute the answer
    • The main agent must be able to defend against criticism or change its opinion
    • This process helps the agent discover its blind spots

Expected Output:

  • A system that, on 50 challenging questions, says “I’m not sure” in 20% of cases (instead of hallucinating)
  • Complete logs of the doubt and reconsideration process for each question
  • Comparative chart: system accuracy with and without the doubt module
  • Analytical report on error patterns in the system (what types of questions cause more mistakes?)

Topic 2: Digital Twin with Deep Persona Modeling

Core Challenge:

Build a digital twin of a real or fictional character that not only has their knowledge but also simulates their thinking style, tone, biases, and even personality evolution over time.

Suggested Capabilities:

  1. Persona Extraction from Text:
    • Extract personality from a dataset (e.g., letters, interviews, books):
      • Linguistic style: sentence length, specific words, grammatical structure
      • Worldview: attitudes toward life, politics, science
      • Biases: cognitive biases (e.g., optimism, pessimism, conservatism)
      • Knowledge: what they know and what they don’t know
  2. Persona-Consistent Response Generation:
    • Each response must be consistent with the persona
    • If it deviates from the persona, a Critic detects and corrects it
  3. Persona Evolution Over Time:
    • Persona is not static! It should evolve based on new experiences
    • For example, if the character encounters new ideas, they might change their opinion
    • Implement an opinion change mechanism
  4. Counterfactual Simulation Mode:
    • “What would this character do if placed in situation X?”
    • Simulate character decisions in various scenarios
  5. Emotional Memory:
    • Each event has an emotional weight (positive/negative)
    • High-emotional-weight events are retained longer in memory
    • This makes the character behave more human-like
  6. Cognitive Dissonance Detection:
    • If the character has two conflicting beliefs, the system detects and shows the conflict
    • The character attempts to resolve the conflict (change belief or justify)

Expected Output:

  • A digital twin of a historical figure (e.g., Albert Einstein or Socrates) or a fictional character
  • A 30-question interview with this character
  • Comparison of character responses at the beginning and end of interaction (persona evolution)
  • Psychological analysis of the generated persona
  • Report: How can this technique be used to simulate customers in marketing?

Topic 3: Automated Scientific Literature Browser

Core Challenge:

Students and researchers spend significant time searching, screening, and reading papers for literature reviews. You need to build a multi-agent system that:

  • Takes a research question in natural language
  • Generates appropriate search queries for arXiv
  • Retrieves and scores papers for relevance
  • Extracts methodology, findings, and future work sections from relevant papers
  • Produces a synthesized report on common themes and research gaps

Proposed Architecture (6 Agents):

  • Query Construction Agent: Converts the research question into multiple arXiv search queries
  • Search Agent: Executes queries and collects results
  • Screening Agent: Scores each paper for relevance to the research question
  • Knowledge Collection Agent: Downloads relevant papers and extracts full text
  • Analysis Agent: Extracts methodology, findings, and future work from each paper
  • Synthesis Agent: Identifies common themes and research gaps, writes the final report

Output: A comprehensive report including paper summaries, main themes, and suggestions for future research

Level 2 Topics

Topic 1: Industrial Data Analysis System with Intelligent Agents

Inspired by research: Agentic Data Analysis for Intelligent Manufacturing

Core Challenge:

Data analysis in manufacturing industries requires technical experts who are expensive and scarce. Non-technical users (production managers, quality controllers) need to be able to ask questions from data and get answers. You need to build a system that:

  • Allows non-technical users to ask questions from industrial data using natural language
  • Analyzes data and generates charts
  • Interprets results and explains them in simple language
  • Operates with high accuracy and interpretability
  • Works for 2 different industrial datasets

Suggested Capabilities:

  1. Lightweight Agentic Framework:
    • Use LangGraph for agent orchestration
    • Agents should be modular to allow adding new tools
  2. Specialized Agents:
    • Query Understanding Agent: Converts user question to data operations
    • Search Agent: Finds relevant data
    • Analysis Agent: Performs statistical analysis and computations
    • Visualization Agent: Generates appropriate charts
    • Interpretation Agent: Explains results in simple language
  3. Evaluation with Industrial Datasets:
    • Use 2 manufacturing datasets (e.g., production line data and quality control data)
    • Design a standard set of questions
    • Compare system accuracy with direct prompting (without agents)
  4. Explainability:
    • Each answer must be accompanied by analysis steps
    • User should see what calculations the system performed and why it reached that conclusion
  5. Integration with Existing Tools:
    • System must work with Pandas, Matplotlib, and Seaborn
    • Accept CSV or Excel data as input

Expected Output:

  • An industrial data analysis system with 5 specialized agents
  • Execution on 2 industrial datasets with 50 standard questions
  • Report: How much better does the agent-based system perform compared to direct prompting?
  • Practical examples of questions and answers

Topic 2: Intelligent Agent for Knowledge Extraction from Scientific Papers and Knowledge Graph Construction

Core Challenge:

Scientific papers contain valuable information (concepts, relationships between concepts, methods, findings), but manual extraction and organization is nearly impossible. You need to build an agent that:

  • Takes a scientific paper as input
  • Extracts key concepts
  • Detects relationships between concepts (e.g., “X causes Y”, “X is a method for Y”)
  • Constructs a knowledge graph from the paper
  • Visualizes the graph

Suggested Capabilities:

  1. Concept Extraction Agent:
    • Using NER (Named Entity Recognition) and key phrase extraction techniques, find main concepts
    • Categorize concepts into “Topic,” “Method,” “Finding,” “Challenge”
  2. Relation Extraction Agent:
    • For each pair of concepts, detect the relationship type:
      • Causes: X causes Y
      • Part-Of: X is part of Y
      • Method-For: X is a method for Y
      • Improves: X improves Y
      • Challenges: X is a challenge for Y
    • Use few-shot learning with standard relationship examples
  3. Graph Builder Agent:
    • Store concepts as nodes and relationships as edges
    • Build the graph with NetworkX or Neo4j
    • Visualize the graph
  4. Graph Summarizer Agent:
    • Produce a textual summary of the graph:
      • Main concepts
      • Most important relationships
      • Conceptual clusters (which concepts are related?)
  5. Paper Comparison Mode:
    • User can compare two papers
    • System builds two knowledge graphs and identifies similarities/differences

Expected Output:

  • A knowledge graph extraction system for scientific papers (with LangChain and multiple Agents)
  • Execution on 20 papers from different fields (e.g., medicine, computer science, physics)
  • 10 visual knowledge graphs with complete analysis

Topic 3: Customer Support Conversation Analysis Agent

Core Challenge:

Companies have massive volumes of customer support conversations that manual analysis cannot handle. You need to build a multi-agent system that:

  • Takes conversations (chat text or call transcripts) as input
  • Summarizes each conversation
  • Detects overall sentiment (positive/neutral/negative)
  • Extracts key topics discussed
  • Generates an analytical dashboard of conversation patterns

Proposed Architecture (4 Agents):

  • Summarizer Agent: Summarizes each conversation into 3-5 key sentences
  • Sentiment Agent: Detects overall tone (positive/neutral/negative)
  • Topic Extraction Agent: Identifies 2-3 main topics discussed
  • Combiner Agent: Aggregates all agent outputs into a cohesive JSON structure

Output: A JSON file containing summaries, sentiment, and topics for each conversation + a Streamlit dashboard

Topic 4: Multi-Agent Simulation of Fictional Societies

Inspired by research: BookWorld

Core Challenge:

A system that turns a novel or story into a living world. Characters behave as independent agents, interact with each other, and create new stories that remain faithful to the original world. But beyond simple simulation, the system must:

  • Extract characters with deep persona from the book
  • Identify and simulate relationships between characters
  • Respect the geography and constraints of the story world
  • Generate new and creative stories in the same world
  • Allow the user to enter the story as a new character

Suggested Capabilities:

  1. World Extraction:
    • From a novel (e.g., Harry Potter, Tolkien, or Persian stories), extract:
      • Characters: traits, goals, relationships
      • World rules: magic, technology, physical constraints
      • Geography: locations, distances, paths
      • History: important events before the story
  2. Dynamic Persona:
    • Each character should evolve based on new experiences
    • Changes must be consistent with the original character
  3. Society Simulation:
    • Characters interact with each other (conversation, cooperation, conflict)
    • Each character has a daily schedule
    • Random events (e.g., wars, disasters) affect society
  4. New Story Generation:
    • User provides an initial spark (e.g., “What if Harry Potter went to a different school?”)
    • System generates a complete story with the same characters and rules
    • Story must be unpredictable yet logical
  5. Counterfactual Mode:
    • “What if character X made a different decision at event Y?”
    • System simulates alternative story branches
  6. User Interaction:
    • User can enter the world as a new character
    • User decisions affect the story’s progression
    • Main characters react to the user (based on their persona)
  7. Source Fidelity:
    • Design an evaluator that compares each output with the original book
    • Calculate a fidelity score (0 to 100)
    • If fidelity is below 80%, reject the output

Expected Output:

  • A simulated world from a famous novel (suggestions: Harry Potter, Lord of the Rings, or Shahnameh)
  • At least 5 new stories generated in the same world (with new events)
  • Analytical report on character evolution across new stories
  • Comparison with traditional methods (story generation with a regular LLM)

Level 3 Topics

Topic 1: Automated Scientific Paper Writing Agent with Research and Citation

Core Challenge:

Writing a scientific paper requires research, literature review, and accurate citation. You need to build an agent that:

  • Takes a scientific topic as input
  • Finds relevant papers from Google Scholar or arXiv
  • Writes a summary of each paper
  • Produces a paper draft with standard structure (Introduction, Methodology, Results, Discussion, Conclusion)
  • Adds accurate citations (in APA or IEEE format)

Suggested Capabilities:

  1. Search Agent:
    • Uses Serper API (Google Search) or arXiv API to find relevant papers
    • Extracts title, abstract, authors, and publication year
    • Ranks papers based on recency and citation count
  2. Study Agent:
    • Writes a summary of each paper (200-300 words) emphasizing: research question, method, findings
    • Extracts keywords and important concepts
    • Identifies research gaps
  3. Writer Agent:
    • Writes the paper based on summaries and standard structure
    • Each section must be supported by sources
    • Scientific language and appropriate tone (formal, precise, objective)
  4. Editor Agent:
    • Reviews the paper for scientific accuracy, coherence, and grammar
    • Provides improvement suggestions
    • Verifies citations (are all sources present in the references?)
  5. Final Output:
    • A LaTeX or Word file with the complete paper
    • Reference list
    • A 200-word executive summary

Expected Output:

  • A scientific paper writing agent with 4 sub-agents (Search, Study, Writer, Editor)
  • Generation of 5 papers in different fields (e.g., machine learning, medicine, economics, social sciences)
  • Report: In which domains does the system perform better?