Irvallen Saragih(he/him)

Computer Science, NLP Engineering, AI/ML Engineering

Major: Computer Science and Communication Engineering (BEng), Waseda University

About

I am a Waseda University graduate, majored in Computer Science and Communication Engineering, with a specialization in Natural Language Processing and strong experience in AI and ML Engineering. My work focuses on building and evaluating machine learning systems, with particular interests in NLP, Large Language Models, and practical AI applications.

My work revolves around building robust AI pipelines and investigating why systems behave the way they do, from diagnosing regression-to-the-mean biases in demographic prediction to benchmarking generative models against sequence taggers using the Appraisal Framework.

When I'm not coding, I typically watch random series or movies (big fan of the MCU and horror genre), getting in my afternoon runs, and writing random journals with my precious fountain pens.

Experience

Undergraduate Researcher

Apr 2025 – Jul 2026

Waseda University, Kobayashi LaboratoryTokyo, Japan

Advisor: Prof. Tetsunori Kobayashi

Engineered NLP text classification models (DA-RoBERTa, PyTorch) on academic discourse datasets, achieving a 0.6973 Macro F1-score and 0.7159 Weighted F1-score on few-shot synthetic data.
Built data augmentation pipelines incorporating 2,000 synthetic text spans per engagement category (Citation, Sources, Endophoric, Justifying) to rebalance training distributions across 11,157 total sentences.
Evaluated transformer-based sequence taggers against baseline architectures via structured offline experimentation to benchmark semantic accuracy.

Backend Developer

Sep 2022 – Jul 2023

Google Developer Student Club (GDSC)Tokyo, Japan

Collaborated in a cross-functional team of 20 members to build and preprocess structured tabular datasets for machine learning challenge submissions.
Wrote data transformation and preprocessing scripts in Python to clean and format training datasets for exploratory data analysis pipelines.
Maintained large codebases utilizing strong Python engineering skills, applying rigorous testing, code review, and CI to ensure complete operational ownership.

Research

Evaluating Data-Centric Augmentation Strategies for Sequence Tagging in Engagement Span Identification

Benchmarking state-of-the-art RoBERTa+LSTM (Eguchi & Kyle, 2023) against data-augmented sequence taggers for discourse-level rhetorical classification.

Computational LinguisticsNLPMachine LearningAlgorithmsData AugmentationspaCyPrompt EngineeringModel Training, Inference & Evaluation

Projects

Semantic MCP Gateway(Most Recent)

An enterprise AI infrastructure control plane that dynamically prunes MCP tool schemas, caches semantically equivalent queries, intercepts adversarial prompt injections, and orchestrates resilient multi-provider LLM failover.

Model Context Protocol (MCP)Custom BM25 / TF-IDF Inverted IndexCharacter 3-Gram & Stemmed L2 Cosine SimilarityGroq APIGoogle GenAI SDK

Multi-Agent Academic Engagement Evaluator

A hybrid NLP pipeline combining a LangGraph multi-agent LLM workflow and a custom Data Augmented RoBERTa (DA-RoBERTa) sequence tagger for automated, discourse-level rhetorical analysis of academic texts.

LangChainLangGraphPydanticUvicornspaCyGroq APIGoogle Cloud PlatformDockerNgrok

Automotive Technical Intelligence RAG System

An optimized, full-stack Retrieval-Augmented Generation (RAG) architecture utilizing hybrid vector search and ONNX-based cross-encoder reranking to deliver highly accurate, document-grounded technical intelligence.

Retrieval-Augmented Generation (RAG)PostgreSQLpgvectorFastAPICI/CDRenderDockersentence-transformersEvaluation

View All Projects

Open Source Software (OSS) Contribution

BerriAI/lite-llm

Case Study: Fixing Audio Transcription Routing for gpt-transcribe in LiteLLM

Status:Successfully Merged

alibaba/open-code-review