Case study 03 / LLM applications · Conversational systems
Smart Financial Assistant
Multilingual dialogue, with deliberate control.
An FSM-controlled LLM dialogue system in English, Bengali and Hindi, with 36 automated evaluation trials.
01 / Context
Problem
A multilingual dialogue system needs to handle language while maintaining a clear conversation flow. Generating a fluent answer and controlling the next dialogue step are separate engineering concerns.
The assistant supports English, Bengali and Hindi, with a finite-state machine controlling the flow around local LLM inference.
02 / Ownership
My Role
Built the multilingual LLM dialogue system with LLaMA 3.1 and Mistral 7B served through Ollama.
Implemented finite-state conversation control and a 36-trial automated evaluation pipeline to inspect dialogue behaviour across the supported languages.
03 / System design
Architecture
- 01User dialogue
English · Bengali · Hindi
- 02State controller
Finite-state conversation flow
- 03Local inference
Ollama · LLaMA 3.1 / Mistral 7B
- 04Evaluation
36 automated trials
04 / Engineering decisions
Engineering Decisions
Separating conversation flow from model output.
The finite-state machine controls conversation flow while the local models provide language generation. This is a confirmed design choice, not a claim about an unreported failure.
TODO: Add a real failure transcript, the state transition or inference change that fixed it, and the corresponding evaluation result.
05 / Evidence
Results
The trial count describes the evaluation pipeline. Pass rate, language-specific scores, latency and model comparisons have not been supplied.
06 / Tools & capabilities
Stack
07 / Explore further
Links
TODO: Replace the profile link with the exact project repository and add reproducible setup instructions.