TBSystemTanmay
Bhuskute
Back to projects
Level 1 Case StudyAI Systems / Jun 2025

Spendora - Conversational AI for Expense Insights

I built an LLM-powered conversational AI assistant that provides personalized expense insights using retrieval-augmented generation and tool-based reasoning.

No published metric yet

01

TL;DR

  • I built an LLM-powered conversational AI assistant that provides personalized expense insights using retrieval-augmented generation and tool-based reasoning.
  • Result metrics are not published yet.

02

Problem

  • How can users get intelligent insights about their spending habits through natural conversation? Anyone wanting AI-powered financial analysis without manual data analysis. Most expense tracking tools lack conversational intelligence and context-aware recommendations.

03

My Role

  • personal build with end-to-end ownership
  • Problem framing: How can users get intelligent insights about their spending habits through natural conversation?
  • Architecture: LangChain conversational agents with memory management
  • Implementation: Custom financial data toolchains, vector embeddings for RAG
  • Evaluation: metrics not published yet
  • Before: Most expense tracking tools lack conversational intelligence and context-aware recommendations
  • Personally designed: Used LangChain for agent orchestration because Simplifies memory management and tool-use patterns for conversational AI; Implemented RAG with custom toolchains because Enables grounding in actual financial data without fine-tuning
  • Others owned: No separate collaborator-owned subsystem is published in the source data.

04

Constraints

  • Built Jun 2025 (personal). Limited time, focus on LLM integration quality.

05

Architecture

  • Input: Natural language queries about expenses
  • Backend: LangChain conversational agents with memory management
  • Data & storage: Custom financial data toolchains, vector embeddings for RAG
  • External APIs: HuggingFace models (Mistral, Gemma), LangChain framework
  • Output: Conversational interface with context-aware financial recommendations
FlowSpendora system flow

LangChain conversational agents with memory management; Custom financial data toolchains, vector embeddings for RAG; Conversational interface with context-aware financial recommendations

  1. input 01Input

    Natural language queries about expenses

  2. process 02Backend
    input ->

    LangChain conversational agents with memory management

  3. storage 03Data / storage
    backend ->

    Custom financial data toolchains, vector embeddings for RAG

  4. external 04External APIs
    backend ->

    HuggingFace models (Mistral, Gemma), LangChain framework

  5. output 05Output
    storage ->external ->

    Conversational interface with context-aware financial recommendations

Routes

  • Input -> Backend
  • Backend -> Data / storage
  • Backend -> External APIs
  • Data / storage -> Output
  • External APIs -> Output

06

Key Technical Decisions

  • Used LangChain for agent orchestration
  • Implemented RAG with custom toolchains

07

Implementation

  • Input layer: Natural language queries about expenses
  • Core system: LangChain conversational agents with memory management
  • Data layer: Custom financial data toolchains, vector embeddings for RAG
  • External boundary: HuggingFace models (Mistral, Gemma), LangChain framework
  • User output: Conversational interface with context-aware financial recommendations

08

What Broke / What Didn't Work

  • Rejected: Building custom conversation loop. Chosen path: Used LangChain for agent orchestration.
  • Rejected: Fine-tuned LLM on expense data. Chosen path: Implemented RAG with custom toolchains.
  • RAG adds latency for retrieval but improves accuracy over hallucination-prone base models
  • Multiple toolchains provide flexibility but increase complexity

09

Results

  • No headline result metric is published yet; this page keeps unsupported claims out of the story.

10

What I'd Change Now

  • Add real-time transaction integration
  • Implement budget forecasting
  • Support multi-currency analysis

11

Stack

  • HuggingFace (Mistral, Gemma models)
  • LangChain
  • Python
  • Vector embeddings

12

Links

  • Source docs: 2-projects.json
Deep dive prompts

Ask me about the trade-offs.

  • Why this architecture boundary exists: LangChain conversational agents with memory management
  • How I evaluated how I would measure the next version
  • The hardest tradeoff: RAG adds latency for retrieval but improves accuracy over hallucination-prone base models
  • What I would change next: Add real-time transaction integration
Contact meMore projects