How I Reduced Inference Latency by 43%
Optimizing a BERT sentiment model from 85ms to 48ms using ONNX Runtime, quantization, and request batching. Includes benchmarks and deployment considerations.
$
█
PRODUCTION
v2.3.1
Shahin.ai
ML Engineer Instance
0
Deployments
0
Training Years
0
Models Built
// about_me.py
Model Architecture
Neural network topology of a production ML engineer
| Layer | Type | Description | Params |
|---|---|---|---|
| INPUT | Foundation | Computer Science & Mathematics | ∞ |
| Dense | Machine Learning Fundamentals | 512 | |
| Conv2D | Deep Learning & Neural Networks | 1024 | |
| Attention | NLP & Computer Vision | 2048 | |
| Transformer | LLMs & Generative AI | 4096 | |
| OPTIM | AdamW | Real-world Deployment & MLOps | lr=0.001 |
| OUTPUT | Softmax | Production-ready ML Systems | ∞ |
Total Parameters
7,680+
Hours of LearningBatch Size
Multi-task
Parallel ProcessingLoss Function
Minimizing
Continuously OptimizingInference Speed
Real-time
Fast Deployment
// experiments.py
Research Experiments & Deployments
Production models shipped with measurable impact
● Deployed
#EXP-001
Dengue Risk Checker — Dhaka
A production-ready Flask web application for predicting dengue risk in Dhaka, Bangladesh using XGBoost, scikit-learn, and advanced feature engineering for individual risk predictions and public health insights.
ML Model
XGBoost
Feature Eng.
Advanced
Coverage
Dhaka-wide
Model: XGBoost + scikit-learn
Backend: Flask (Production-ready)
Features: Risk prediction + Public health insights
● Deployed
#EXP-002
ResearchHub — ML Paper Manager
A lightweight Flutter mobile app for research paper discovery and management with built-in ML capabilities for automated categorization, clustering, semantic search, and smart recommendations.
Categorization
Auto
Search
Semantic
Clustering
Topic-based
Frontend: Flutter (Mobile)
ML Pipeline: Embeddings + Clustering
Mode: Online / Offline-friendly
// training_logs.py
Training History
Model training logs — from initialization to production
[2019]
Model initialized
— BSc in Computer Science & Engineering
INIT
[2020]
Feature engineering
— Python, Statistics, Data Analysis mastered
TRAINING
[2021]
Gradient boost applied
— ML Specialization certified, first models built
TRAINING
[2022]
Loss decreasing rapidly
— First freelance ML deployment shipped
DEPLOYED
[2023]
Production model shipped
— Deep Learning, NLP & CV projects in production
DEPLOYED
[2024]
Scaling phase initiated
— MLOps, Cloud integration, CI/CD pipelines
SCALING
[2025]
LLM fine-tuning epoch
— Generative AI, RAG systems, LLM deployment
SCALING
[2026]
Global convergence achieved
— Full-stack ML Engineer, production-ready
LIVE
[NEXT]
Continuous learning...
— Always updating weights
▶ RUN
Training & Validation Loss
// live_demos.py
Live ML Demos
Interactive machine learning models running in your browser — proof that I deploy, not just code.
Sentiment Analyzer
LIVEType anything → Get real-time sentiment prediction with confidence scores.
Awaiting input...
Ask My Portfolio
LIVEChat with my portfolio AI — ask about skills, projects, or experience.
Resume Classifier
LIVEPaste resume text → Predicts if the candidate fits an ML Engineer role.
Awaiting input...
House Price Predictor
LIVEInput house features → Get instant price prediction from trained model.
Awaiting input...
// writing.py
Writing & Research Notes
Lessons learned from building and deploying ML systems
Deploying ML Models at Scale with Docker
From single-container deployments to multi-service architectures. Docker Compose for ML, health checks, resource limits, and graceful shutdown patterns.
When NOT to Use Deep Learning
A decision framework for choosing between traditional ML and deep learning based on data, interpretability, latency, and maintenance overhead.
Building RAG Systems That Actually Work
Chunking strategies, embedding model selection, re-ranking, and evaluation metrics that correlate with user satisfaction in production RAG pipelines.
The Gap Between Jupyter and Production
What I learned shipping my first 5 ML models. Error handling, input validation, logging, model versioning, and the unglamorous work that makes systems reliable.
Feature Store Patterns for Small Teams
You don't need Feast or Tecton to start. Practical patterns for feature pipelines with DVC, SQLite, and simple Python scripts that scale.
// lets_connect.py
Let's Build Something Together
Got a problem worth solving? I turn messy data into production systems. Reach out — I reply fast.
GitHub
github.com/shahin5646 View repositories →$ whoami
Shahin — ML Engineer
$ cat ./connect.json
{
"name": "Shahin",
"role": "ML Engineer",
"email": "shahinok1912@gmail.com",
"github": "github.com/shahin5646",
"linkedin": "linkedin.com/in/shahin-5646-diu",
"status": "Open to opportunities",
"response_time": "< 24 hours",
"interests": [
"Production ML Systems",
"NLP & LLM Applications",
"Computer Vision Pipelines",
"MLOps & Infrastructure"
]
}
$ echo "Let's collaborate!"
Let's collaborate!
Send Me a Message
Have an idea, a project, or just want to say hi? Drop a note below.