/ Projects

Things I've built.

RAG & LLM Benchmarking - English and Nepali

This experiment benchmarked multiple chunking strategies for financial-report RAG and found that section-aware chunking produced the strongest overall answer quality, grounding, and numeric reliability.

2026

Workflow:

Financial PDFs were parsed into layout-aware blocks using page text, headings, font signals, tables, and paragraph structure before chunking.
Chunking strategies tested included recursive 512/768/1024, paragraph-hybrid 512/768, section-aware 512, table-aware paragraph-hybrid 512, and a parent-child design.
Each strategy used the same embedder, generator, retriever, prompt, judge, and top-k retrieval setup to keep the comparison controlled.
Recursive chunking achieved perfect retrieval coverage, with Hit@K and Evidence Recall both reaching 1.0000 across recursive variants.
Section-aware chunking performed best across answer-level metrics: F1 0.4721, Numeric Accuracy 0.4667, Citation Faithfulness 0.7556, Judge Score 7.6667, Judge Correct 0.6000, Judge Faithful 0.8000, and Judge Used Evidence 0.8000.

AI for Business Insights - Text-to-SQL Analytics System

Production-ready text-to-SQL system that converts natural-language questions into validated SQL and business insights.

2025

Workflow:

Structured business data stored in PostgreSQL
Query understanding to identify relevant tables and fields
Top-k table retrieval with fuzzy matching over column values
SQL generation and execution for analytical datasets
Python sandbox execution for chart generation

AI Sales Copilot for B2B Workflows

AI assistant for automating B2B sales communication. Generates personalized outreach, summarizes leads, and prepares follow-ups using CRM context and company data.

2025

Workflow:

Lead and CRM data ingestion from internal sources
Embedding generation and storage in Qdrant
Semantic retrieval of company context and past interactions
RAG pipeline for personalized email generation
Conversation summarization for lead tracking

Text-to-Reels AI Content Generator

Generative AI system for creating short-form videos from text prompts by combining gameplay footage, generated visuals, and AI voiceovers.

2025

Workflow:

Gameplay video template selection for the lower screen region
LLM-generated narration and content plan from user text
Image generation for upper-screen visuals
ElevenLabs voiceover generation
Automated video assembly into final reel output

Nepali Cultural Video Understanding

Vision-language system for captioning and QA on Nepali cultural videos. Multimodal transformers align visual features with Nepali language semantics.

2024

Workflow:

Scraped Nepali video data from YouTube
Used capable VLMs to generate captions + QA
Frame processing with 10 second clips
VLM and encoder-decoder architecture
Generated final captions and QA outputs

Stock Market Volatility Forecasting

Predictive system for estimating market volatility and Value-at-Risk using historical S&P 500 data with hybrid GARCH-LSTM architectures.

2024

Workflow:

Collected and preprocessed historical stock data from Yahoo Finance
Modeled volatility using GARCH and related variants
Integrated hybrid GARCH-LSTM models for short-term forecasting
Backtested predictions for risk estimation accuracy
Produced Value-at-Risk estimates and decision-support visualizations

Sign Language Recognition with Graph Neural Networks

System for recognizing sign language gestures from video using graph neural networks and temporal sequence modeling.

2022

Workflow:

Extracted hand and body keypoints from video frames
Constructed spatio-temporal graphs for joints and gestures
Applied Graph Attention Networks for spatial dependencies
Integrated LSTM layers for temporal sequence modeling
Predicted sign language gestures on benchmark datasets