Hi, this is

S.S. Tarek

I build

AI/ML Engineer turning complex problems into production-grade intelligent systems.

View My Work Get In Touch
S.S. Tarek

About Me

Hi. I'm an AI/ML and AIOps Engineer, experienced in both R&D and production-grade intelligent systems: from data pipelines and model training to agentic AI and cloud-native deployment.

My work sits at the intersection of research and engineering. I prefer to design systems that work in production and bring value to real-world problems. Whether that's an autonomous incident investigation pipeline that posts root cause analysis directly to Slack, or a music analysis model deployed inside a Sony library, I focus on the full journey from idea to impact.

Outside the engineering life, my horizon spans across books, anime, and most predominantly, music. A shade of Rock n Roll keeps my soul afloat and acts as the driving force behind my everyday life.

Tech Stack

Languages

Python Python
SQL SQL
C C

ML & Deep Learning

PyTorch PyTorch
TensorFlow TensorFlow
Keras Keras
scikit-learn scikit-learn
XGBoost XGBoost

Generative AI & NLP

LangChain LangChain
LangGraph LangGraph
Hugging Face Hugging Face
BERT BERT
MCP MCP

Cloud & Infrastructure

AWS AWS
Docker Docker
Kubernetes Kubernetes
Terraform Terraform
GitHub Actions GitHub Actions

MLOps & APIs

MLflow MLflow
FastAPI FastAPI
OpenTelemetry OpenTelemetry
Prometheus Prometheus
Grafana Grafana

Data Engineering

NumPy NumPy
Pandas Pandas
PostgreSQL PostgreSQL
Redis Redis
Dask Dask
PySpark PySpark

Experience

Woven by Toyota

via Robert Half AI/ML & AIOps Engineer
Feb 2026 - Jun 2026 Tokyo, Japan

Designed and deployed production-grade AIOps systems on AWS - owning the AI roadmap for cloud and network operations, replacing manual workflows with autonomous AI pipelines.

  • Built and deployed an event-driven RCA pipeline that automatically investigates AWS incidents reported in Slack and posts root cause analysis in real time
  • Extended the RCA pipeline into a fully agentic system using LangChain and LangGraph with dynamic MCP tool binding, implemented as two microservices containerized with Docker and orchestrated with Kubernetes
  • Deployed an AI alert classification system achieving 97% accuracy and 100% recall on Critical alerts, replacing manual review of ~30 daily alerts
  • Designed an SLO violation forecasting pipeline using Holt-Winters and Monte Carlo confidence scoring - average violation lead-time of 19 days on validation data, against a target of 7 days
Python AWS LangChain LangGraph MCP Docker Kubernetes Terraform FastAPI ChromaDB GitHub Actions

Sony Computer Science Laboratories

via Hiperdyne Corporation AI Engineer - Deep12 Music Analysis AI
Nov 2018 - Apr 2025 Tokyo, Japan

R&D contributor on Deep12, a music analysis AI developed in collaboration with Sony Computer Science Laboratories and deployed as a service in the Sony Music Publishing library.

  • Designed and trained CNN, LSTM, and Transformer-based models for music search, detection, and prediction tasks - achieving 83-89% accuracy
  • Built a multi-stage NLP pipeline progressing from TF-IDF & Naive Bayes to fine-tuned BERT, improving performance by ~8% over the next-best model
  • Leveraged AWS Bedrock for synthetic minority-class generation using few-shot prompting with deduplication and topic filtering at scale
  • Optimized GPU-accelerated data processing pipelines with PyTorch parallelization - improving runtime by 4-5x
Python PyTorch TensorFlow Keras scikit-learn BERT AWS MLflow Docker GitHub Actions

Projects

Case Study

AI-Driven Root Cause Analysis

When an AWS incident report drops in Slack, the system investigates autonomously - with a root cause delivered in the same thread.

  • Multi-step Lambda pipeline orchestrated by Step Functions - validates, retrieves CloudWatch logs, analyzes with LLM, and posts root cause report to the originating Slack thread
  • Extended into a fully agentic system with LangChain and LangGraph - stateful orchestration, dynamic MCP tool binding, prompt injection safeguards at the MCP layer
  • Implemented as two microservices - FastAPI inference layer and custom MCP tool server, containerized with Docker and orchestrated with Kubernetes
  • TF-IDF log deduplication reduced LLM input tokens by up to 96% with no loss in analysis quality
Python AWS LangChain LangGraph MCP FastAPI Docker Kubernetes Step Functions
Case Study

AI-Driven Alert Classification

Replaced manual review of ~30 daily Slack alerts with an automated triage pipeline.

  • Tiered classification architecture - nearest-neighbor retrieval, regex pattern matching, RAG-based LLM escalation with confidence scoring and a fail-safe default to Critical
  • 97% accuracy, 0.89 macro F1, 100% recall on Critical alerts
  • Deployed on AWS Lambda with Terraform-provisioned infrastructure and GitHub Actions CI/CD
  • CloudWatch error-rate alarms with SNS notifications for production pipeline health monitoring
Python AWS Lambda Bedrock ChromaDB OpenSearch Terraform GitHub Actions
Case Study

Deep12 - Music Analysis AI

R&D contributor on a music analysis AI deployed inside the Sony Music Publishing library.

  • Designed and trained CNN, LSTM, and Transformer-based models for audio-based music classification, similarity search, and attribute detection - achieved 83-89% accuracy
  • Built a multi-stage NLP pipeline from TF-IDF & Naive Bayes to fine-tuned BERT - 8% improvement over next-best model
  • Synthetic minority-class generation via AWS Bedrock with few-shot prompting, cosine similarity deduplication, and BERTopic topic filtering
  • GPU-accelerated PyTorch data processing pipelines - 4-5x runtime improvement
Python PyTorch TensorFlow BERT AWS Bedrock MLflow Docker GitHub Actions
Personal Project

LLM Uncertainty Quantification

Most LLMs tell you what they think. This project asks how confident they really are.

  • Measured how reliably a model's confidence predicts correctness across math reasoning and factual QA tasks - using 7 signals including token entropy, sequence confidence, and log-prob variance
  • AUROC of 0.75 on reasoning tasks and 0.84 on factual QA - uncertainty signals rank errors well above random across both task types
  • Identified confident ignorance - the model can be highly certain and completely wrong, especially on knowledge-recall tasks
  • Reproduced entirely on a single consumer GPU using a 4-bit quantized model - no cloud compute required
Python PyTorch Hugging Face Phi-3-Mini GSM8K TriviaQA
View on GitHub
Personal Project

LLM Fine-tuning with LoRA

Fine-tuned a 3.8B parameter LLM to produce structured JSON output reliably - end-to-end, from synthetic data generation to adapter training on AWS EC2.

  • Phi-3 Mini (3.8B) fine-tuned with LoRA adapters applied to 4 attention and feedforward modules - 25M trainable parameters, trained on AWS EC2 G4dn instance
  • Synthetic training labels generated via AWS Bedrock with the existing emotion label as an anchor; 1,174 valid examples after filtering with stratified train/val/test splits
  • Schema compliance improved from 4.6% strict baseline to 100% after fine-tuning: overall accuracy rising from 54% to 81%
  • EC2 G4dn instance fully provisioned via Terraform - VPC, IAM role, and S3 access policy as code
Python PyTorch Hugging Face PEFT LoRA AWS EC2 Bedrock Terraform W&B Docker
View on GitHub
Personal Project

PDF RAG API

Upload a PDF. Ask questions. Get grounded answers with exact page citations.

  • Async ingestion pipeline - long-running processing happens in the background after the request returns, surviving container restarts
  • Page-level chunking preserves page boundaries enabling exact citation of source pages per answer
  • Distributed tracing pinpointed the embedding step as the ingestion bottleneck across concurrent requests - not derivable from logs alone
  • 27 passing tests with async route testing and mocked AWS dependencies, CI on every push
Python FastAPI AWS Bedrock S3 Vectors LangChain OpenTelemetry Terraform Docker
View on GitHub

Contact

Let's build something.

I'm currently open to new opportunities - whether that's a full-time role, contract work, or an interesting conversation. My inbox is open.

sstarek93@gmail.com Download Resume
linkedin.com/in/sstarek HeavyMetalAlchemist Yokohama, Japan