← Back to work

Generative AI · Conversational Systems

LLM-Based Financial Assistant Prototype

An individual conversational-AI prototype exploring prompt structure, lightweight model comparison, synthetic financial profiles, and a Gradio interface.

Individual coursework prototype · clearly scoped

01

Documents

02

Embeddings

03

Retrieval

04

Response

Conversational prototype workflow

The public scope reflects the executed notebook rather than claiming an unverified production system.

Scope

Role and problem

My role: Built the notebook prototype, structured prompting flow, lightweight model comparison, and Gradio interface.

Financial-assistant interfaces need explicit scope boundaries. This prototype explores how structured prompts and a simple dialogue interface can organise responses without presenting the result as professional financial advice.

Architecture

System flow

01

Synthetic financial profiles

02

Structured prompt template

03

GPT-2 and DistilGPT-2 comparison

04

Dialogue wrapper

05

Gradio interface

06

Scoped prototype output

Evidence

Measured signals

2 models

Lightweight comparison

GPT-2 and DistilGPT-2 pipelines are wired into the notebook.

Gradio

Interactive interface

A user-facing prototype accepts age, income, risk, goal, and model choice.

Synthetic

Data boundary

The notebook uses synthetic profiles and is not financial advice.

Contribution

  • Built the prompt-driven dialogue prototype and Gradio interface.
  • Compared lightweight model pipelines inside one notebook workflow.
  • Keep the public scope narrow: conversational prototype, synthetic profiles, and no financial-advice claim.

Lessons

  • Interface clarity and scope boundaries matter as much as model output.
  • Synthetic examples are useful for prototyping but do not establish real-world financial suitability.
  • A stronger future version would need grounded retrieval, domain evaluation, and safety review.

Limitations

  • The current notebook is a lightweight prototype using synthetic profiles.
  • The public case study does not claim validated financial recommendations.
  • Grounded retrieval and multilingual evaluation are future extensions rather than published evidence.

Stack

  • Python
  • Transformers
  • GPT-2
  • DistilGPT-2
  • Prompt Design
  • Gradio