TagIt Project

---

TAGIT

████████╗ █████╗  ██████╗ ██╗████████╗
╚══██╔══╝██╔══██╗██╔════╝ ██║╚══██╔══╝
   ██║   ███████║██║  ███╗██║   ██║   
   ██║   ██╔══██║██║   ██║██║   ██║   
   ██║   ██║  ██║╚██████╔╝██║   ██║   
   ╚═╝   ╚═╝  ╚═╝ ╚═════╝ ╚═╝   ╚═╝   

      TAGIT — Smart Labels for Smart Money

---

TAGIT — AI Transaction Categorization System

Smart Labels for Smart Money

A hybrid AI system that classifies financial transactions using TF‑IDF + Logistic Regression, DistilBERT Transformers, and a clean Streamlit UI.

---

Overview

TAGIT intelligently categorizes messy transaction strings like:

"UPI/ROHAN@OKHDFC/9823"
"AMZN MUMBAI 4093"
"POS 42342 CAFE COFFEE DAY"
"ZOMATO*ONLINE ORDER"
"HPCL/FUEL/PUNE"

It uses a two‑stage hybrid pipeline:

• Baseline Model (Fast): TF‑IDF + Logistic Regression
• Transformer Model (Accurate): DistilBERT + Tabular Features
• Hybrid Router: If baseline is confident → use baseline, else fallback to powerful Transformer

TAGIT also includes a sleek Streamlit interface for real-time testing and CSV batch predictions.

---

Architecture Diagram

                   ┌────────────────────────────┐
                   │         RAW INPUT           │
                   │  (UPI / POS / CARD / etc.)  │
                   └────────────────────────────┘
                                 │
                                 ▼
                    ┌────────────────────────────┐
                    │       PREPROCESSOR         │
                    │ Clean text, numbers, dates │
                    │ Extract merchant token     │
                    └────────────────────────────┘
                                 │
          ┌──────────────────────┼──────────────────────┐
          │                      │                      │
          ▼                      ▼                      ▼
┌───────────────────┐   ┌────────────────────┐   ┌────────────────────┐
│  BASELINE MODEL    │   │ TRANSFORMER MODEL │   │   RULE ENGINE      │
│   TF-IDF + LR      │   │ DistilBERT Hybrid │   │ (optional)         │
└───────────────────┘   └────────────────────┘   └────────────────────┘
          │                      │                      │
          └──────────────┬──────┴──────────────┬───────┘
                         ▼                     ▼
                   ┌────────────────────────────────────┐
                   │        TAGIT HYBRID ENGINE         │
                   │ Baseline if conf ≥ 0.70            │
                   │ Else Transformer                    │
                   └────────────────────────────────────┘
                                 ▼
                    ┌────────────────────────────┐
                    │       FINAL CATEGORY        │
                    └────────────────────────────┘

---

Installation

1️ Create virtual environment

python -m venv .venv
source .venv/bin/activate     # macOS/Linux
.venv\Scripts\activate      # Windows

2️ Install dependencies

pip install -r requirements.txt

---

Usage

Generate synthetic Data

python generate_synthetic.py

---

Training the Transformer Model (DistilBERT + Tabular Features)

TAGIT uses a hybrid Transformer architecture that merges DistilBERT embeddings with numeric features (amount, amount_bucket, weekday, month) for superior classification accuracy.

---

1. Prepare Preprocessed Data

python preprocess.py data/transactions.csv data/preprocessed.csv

This generates:

merchant_clean
merchant_token
amount
amount_bucket
weekday
month
label

---

2. Train the Transformer Model

Run:

python train_transformer.py

This script will:

• Load preprocessed data
• Tokenize merchant text using DistilBERT
• Train hybrid encoder (Transformer + Tabular MLP)
• Save all required model files

Saved Artifacts

File	Purpose
models/transformer_best.pt	Best model weights
models/transformer_label_encoder.joblib	Encodes label strings
models/transformer_scaler.joblib	Scales numeric features
models/tokenizer/	DistilBERT tokenizer
models/transformer_metadata.joblib	Model metadata

---

3. Predict Using Transformer

python predict_transformer.py

---

4. Hybrid Mode (Baseline + Transformer)

python smart_predict.py

Logic:

if baseline_confidence >= 0.70:
    use baseline
else:
    use transformer

Results saved to:

data/predictions_hybrid.csv

---

5. Evaluate Transformer

python eval.py

Outputs macro/weighted F1 and per‑label metrics.

---

GPU Acceleration (Optional but recommended)

Install CUDA‑enabled torch:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Verify:

import torch
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0))

---

Train Baseline Model

python train_baseline.py

Produces:

models/baseline_pipe.joblib
models/label_encoder.joblib

---

Predict (Baseline)

python predict.py

---

Evaluate

python eval.py

---

Streamlit App (TAGIT Dashboard)

streamlit run app2.py

Visit:
http://localhost:8501

---

Highlights

• Real-time baseline inference
• High-accuracy Transformer model
• Smart hybrid confidence routing
• Beautiful Streamlit dashboard
• Clean architecture & modular design
• Easy to extend
• Professional metrics (macro/weighted F1)

---

Team Diamonds

Made with ❤️ for innovation.