TECHNICAL CASE STUDY

BHASABODH — AI Translation System & QA Chatbot

Fine-Tuning Transformer LLMs for Low-Resource Romanized Nepali

Executive Overview

BHASABODH is a bidirectional translation engine and QA assistant designed to convert informal Romanized Nepali text into formal Devanagari script. Built during my internship at eSewa, it leverages fine-tuned BLOOM and BERT models served via a high-performance FastAPI interface.

The Problem

Digital wallet users in Nepal frequently type transaction details, queries, and descriptions in Romanized Nepali (Latin alphabet), whereas backend systems and localization protocols require Devanagari script. Off-the-shelf translation engines struggle with colloquial spelling variants and low-resource language nuances.

My Engineering Role

AI/ML Engineering Intern — Responsible for NLP data preprocessing pipelines, transformer fine-tuning experiments, BLEU metric evaluation, and FastAPI web integration.

System Architecture Flow

[01]Raw Romanized Text Input → Preprocessing & Tokenization Pipeline

[02]Fine-Tuned BLOOM / BERT Variant Model → Devanagari Sequence Generation

[03]BLEU Metric Validation & Post-Processing Guardrails

[04]FastAPI Async Server Layer → Tailwind CSS Frontend Interface

Key Architectural Decisions

Selected BLOOM & BERT variants over generic translation APIs

Generic APIs failed to parse localized Nepali slang and phonetic typing conventions; custom fine-tuning provided exact domain alignment.

FastAPI for Model Backend

FastAPI’s native async support enabled non-blocking model inference requests compared to traditional synchronous WSGI frameworks.

Verified Outcomes & Results

✓Successfully produced a high-accuracy bidirectional translation model tailored to digital wallet interaction contexts.
✓Reduced translation error rate on key transaction terms.
✓Delivered a functional FastAPI prototype ready for internal R&D evaluation.

Lessons Learned

•Data quality and preprocessing pipeline design have a far greater impact on low-resource NLP performance than model size alone.
•Evaluation metrics like BLEU must be paired with human qualitative review for nuanced language tasks.