Hriday Saha
EXPERIENCE
4+ years in production ML across insurance, banking and compliance
FOCUS
Fraud and anomaly detection · NLP and retrieval ·Forecasting · MLOps
CURRENTLY
Data Science Trainer and Consultant, Bengaluru, India
OPEN TO
Global relocation · visa sponsorship welcome · available now
CONTACT
MEASURED IMPACT
Fraud prevented, audit hours removed, reporting automated.
ON-TIME DELIVERY
Across a $1.2M analytics portfolio.
COUNTRIES WORKED IN
Malaysia, France, India.
ROLES HELD
Associate to Senior, then independent.
01
The decision comes first
Before a feature is engineered I want to know who acts on the output and when. A score nobody is allowed to act on is a report with extra steps. That framing is why the uplift work finished as a spend policy rather than a benchmark placing, and why the audit model arrived with a review queue attached to it.
Ends in a decision, not a leaderboard place
02
Constraints before models
Latency budget, retraining window, who signs it off, what the data actually looks like on a bad day. These decide the architecture far more than the choice of algorithm does, and they are cheapest to discover before anything is built rather than during handover.
A sub-100ms budget, fixed before the first model
03
Ship it, then watch it
Launch-day accuracy is the easy half. Populations move, upstream schemas change, and nothing throws an error. The model just quietly gets worse. Monitoring and a retraining plan are part of the design conversation, agreed with engineering and DevOps before release, not bolted on after the first complaint.
Drift checks and a retraining strategy, defined pre-launch
04
Explain it, or it will not be adopted
An auditor will not accept a number with no reason attached, and they are right not to. Confidence scores and key-phrase highlights were not decoration on the mapping system. They were the reason multiple stakeholder groups let it into their process at all.
384 hours a quarter handed back, and auditors could see why
05
Bring people with you
Most of what stops a model reaching production is not technical. It is a risk owner who was not consulted, a compliance step nobody scheduled, an analyst who was handed a dashboard instead of being taught the thing. I plan for that time, because it is the work, not an interruption to it.
3–5 person team · 40+ stakeholders · 98% on-time
MACHINE LEARNING & AI
Supervised
Unsupervised
XGBoost
LightGBM
Anomaly Detection
Transformers
SpaCy
Deep Learning
Sentence-BERT
PROGRAMMING & ML STACK
Python
SQL
R
scikit-learn
TensorFlow
PyTorch
Hugging Face
DATA SCIENCE & ANALYTICS
Predictive modelling
Time-series forecasting
A/B testing
Hypothesis testing
Feature engineering
VISUALISATION & BI
Power BI
Tableau
Matplotlib
Plotly
Seaborn
Dashboard design
INFRASTRUCTURE & MONITORING
Python ETL
Evaluation frameworks
Validation pipelines
Post-deployment monitoring
PSI drift detection
Supervised
Supervised
DEPLOYMENT & MLOPS
Docker
FastAPI
Containerisation
GitHub
Git
API scoring services
COLLABORATION & STRATEGY
Stakeholder management
Technical mentoring
ML-to-business translation
EDUCATION
MSc, Data Science and Business Analytics
2021
Asia Pacific University of Technology and Innovation, Malaysia
BTech, Computer Science and Engineering
2018
MCKV Institute of Engineering, India
CERTIFICATIONS
Microsoft Power BI Analytics
2022
SAS Data Science Professional
2021
IBM Data Science Professional Certificate
2020
Supervised
Supervised
AWARDS AND RECOGNITION
AXA Innovation Excellence Award
2023
An enterprise-grade ML automation system delivering $180K+ in annual savings.
AWS Build Malaysia, Semifinalist
2020
An advanced ML solution in cloud-based data processing. Top 10% of 500+ participants.
Best Rotaract President, RID-3291
2018
Top district president among 60+ clubs, leading 80+ events, co-organising 25+, speaking at 3.