Digvijay Singh Shekhawat

FIFS Gameathon 2025 2.0 · 1st Place

IPL fantasy-point prediction with PyCaret and PuLP integer programming for team optimization, deployed as a Streamlit app with API-based data ingestion and CSV export.

• Won 1st Place at FIFS Sports Data Gameathon 2025 2.0.
• Engineered player-form, venue, and matchup features using pandas and NumPy for IPL fantasy-point prediction with PyCaret models.
• Implemented PuLP integer programming to optimize 11-player teams within 100 credits and 15-player squads within 145 credits; deployed a Streamlit app with API-based data ingestion and CSV export.

Section routing · compilation repair

LLM-integrated LaTeX editing pipeline with section-level request routing, PDF compilation, and human-in-the-loop change approval.

• Routes editing instructions to specific resume sections to preserve unrelated LaTeX content.
• Compiles candidate PDFs and exposes proposed changes for human-in-the-loop review.
• Integrates Gemini, Groq, and Mistral APIs for contextual edits and compilation-error repair.

Source ingestion · grounded generation

Multi-format source ingestion and evidence-grounded LLM generation for campaign planning, with review-state tracking, scheduling, and structured exports.

• Ingests PDF, DOCX, XLSX, and CSV content into a unified planning workflow.
• Validates generated ideas against source evidence and tracks approval decisions before scheduling.
• Persists reusable topic libraries and portable workspaces with Markdown and CSV serialization.

OCR · classification · extraction

Tesseract OCR and LLM-assisted information extraction pipeline for document classification, schema-oriented field extraction, and structured JSON output.

• Extracts text from scanned images and PDFs with Tesseract OCR.
• Integrates Gemini and Groq for document classification and structured field extraction.
• Exposes extracted attributes, document summaries, and serializable JSON results through Streamlit.

AST validation · configurable 5-second timeout

Streamlit analytics assistant that turns natural-language queries into pandas code for CSV and Excel analysis, with interactive Plotly visualizations and contextual follow-up queries.

• Built a Streamlit analytics assistant that converts natural-language queries into pandas code for CSV/Excel analysis and interactive Plotly visualizations.
• Implemented dataset profiling, AST-based code validation, and subprocess execution with a configurable 5-second timeout.
• Integrated Groq, Mistral, and Gemini APIs for LLM routing, provider fallback, and conversation summarization to support contextual follow-up queries.

BERT · 0.8830 weighted F1

LIAR2 classification pipeline combining claim text, fact-checker justifications, and speaker credibility, with a fine-tuned BERT hybrid classifier achieving 0.8830 weighted F1 versus 0.8684 for SVM.

• Developed a LIAR2 classification pipeline using claim text, fact-checker justifications, and speaker credibility, with 18,369 training and 2,296 test examples.
• Fine-tuned bert-base-uncased with a hybrid text–credibility classifier in PyTorch, achieving 0.8830 weighted F1 versus 0.8684 for SVM.
• Benchmarked SVM, Random Forest, and BiLSTM; engineered 50,000 TF-IDF features and applied SHAP to explain SVM predictions across 200 test samples.

Patch embeddings · self-attention

Custom PyTorch Vision Transformer for binary chest X-ray classification using patch embeddings, learned positional encodings, and multi-head self-attention.

• Implements pre-LayerNorm transformer blocks with residual connections and a GELU-based MLP classification head.
• Trains with AdamW and cross-entropy loss, validation-F1 checkpointing, and early stopping.
• Exposes image-upload inference through a Streamlit interface backed by the trained PyTorch model.