Ittoop Shinu Shibu
Applied ML Engineer · Data Scientist
About
I enjoy the whole process of building AI systems: understanding the problem, working with data and models, getting it into use, and improving it after launch. That last part is where the interesting questions are, like whether it actually helps the people using it.
I recently finished my MS in Computer Science at Boston University. Right now I'm working on how reliably LLM judges agree with human evaluators, and how to measure whether an LLM system is useful to the people it serves.
Before that, five years at Singtel, Singapore's largest telecom, owning propensity models on a $2M BCG engagement and designing the holdout that measured their impact. At BU I led technical delivery of four client ML projects, including a retrieval system for the Boston Public Library, and co-authored a paper on deepfake detection and courtroom admissibility.
Right now, I am drawn to LLM and agent evaluation, retrieval and search, inference engineering, and red-teaming.
Experience
Data Scientist
Singtel, Singapore
XGBoost · LightGBM · Llama 4 RAG · Databricks · Spark · MLflow · Snowflake clean room
- Built and deployed propensity and customer-product alignment models on Singtel and Etiqa insurance data in a Snowflake clean room during a $2M BCG engagement: +80% YoY lead conversion, +210% product-line revenue growth.
- Designed the rollout measurement: A/B tests against a random 10% no-touch holdout held for three months; lifted cross-sell conversion +10% MoM.
- Piloted a Llama 4 RAG system for analyst reporting: 86% agreement with analysts, 50% less reporting time.
- Ran models as Databricks jobs (Spark, MLflow) with scheduled retraining, drift monitoring, and pre-release holdout validation.
Business Analyst, promoted to Manager
Singtel, Singapore
Survival models · regularized regression · Python · R · Airflow · Hadoop
- Built a customer lifetime value model in Python and R that improved segment profitability by 15%.
- Maintained Airflow pipelines on a Hadoop cluster, cutting delivery timelines ~20%.
- Managed four analysts, raising team productivity by 30%.
Special Initiatives Intern
BU Spark!, Boston University
- Scoped 6+ client ML projects; prototyped an AI Chinese-character learning app and an IATI data pipeline (Python ETL, Neo4j).
Graduate Research Assistant, Deepfake Forensics
Boston University
- Built frame-extraction and dataset-standardization pipelines for DF40, FaceForensics++, Celeb-DF v2, and DeepSpeak v2, and evaluated detection models on A40 and H100 GPUs.
- Co-author on a deepfake-forensics paper accepted at IDFC 2026; see Publications.
ML Technical Project Manager
BU Spark!, Boston University
- Led technical delivery of four client ML projects with teams of 4 to 6, including the Boston Public Library retrieval system below.
Education
MS, Computer Science
Boston University · GPA 3.67
BS, Data Science
University of California, San Diego · Provost Honors
Selected Work
LLM-as-Judge Reliability Audit
Python · Anthropic and OpenAI judges · Cohen's κ · bootstrap CIs
Pre-registered replication on MT-Bench: across three judges and 100 pairwise comparisons, judge–human agreement ran 0.17 to 0.21 below the human–human baseline, and judges picked ties on 3 to 9% of pairs versus 23 to 25% for humans.
Hybrid Retrieval for the Boston Public Library
BGE-M3 dense and sparse in pgvector · Neo4j entity graph · RRF + metadata rerank · GraphRAG expansion · GPT-4o intent parsing and cited generation · Streamlit on Hugging Face Spaces
Led technical delivery of hybrid retrieval with cited answers over library collections: MRR 0.049 to 0.209 and Hit@10 0.029 to 0.429 on 35 held-out queries.
Generative Augmentation for Hand Osteoarthritis
CycleGAN · WGAN-GP · conditional VAE · ResNet-18 · 41,051 radiographs
Tested three generators at five augmentation ratios under a 135:1 class imbalance; none improved KL4 recall.
Wearable IMU Exercise Classification
PyTorch · CNN-Transformer · 164 Apple Watch recordings
Benchmarked five architectures with recording-level cross-validation; CNN-Transformer reached 90.0% ± 5.7% accuracy.
Publications
Detecting Deepfakes for the Courtroom: A Multi-Modal Forensic Output Framework
Zulal Akarsu, Maryan Rizinski, Ming Zhang, Kyung-Shick Choi, Deven Shah, Ittoop Shinu Shibu, Chan Woo Shin, Lou Chitkushev. International Conference on Intelligent Digital Forensics and Cybersecurity (IDFC 2026). Accepted.
Awards
Winner, Civic Hacks 2026
MLH Best Use of Gemini API and infrastructure track, for ForTheCity, sidewalk accessibility validation
4th place, BU DS+X Hackathon
For CollabNet, research collaborator discovery over OpenAlex
Singtel Spot Award
Insurance Business
Singtel Star Award
Customer Success
Skills
Python · PyTorch · scikit-learn · LightGBM · SQL · Spark · Databricks · Snowflake · MLflow · Airflow · Docker · RAG · pgvector · Neo4j · LangGraph