$ loading ml_portfolio.py ...
M.S. Data Science — WPI '25

Hi, I'm Sheroz
Shaikh

ML Engineer & Data Scientist building production AI systems that deliver measurable business impact. From deploying LLM-powered medical coding systems serving 10+ enterprise clients to building semantic search platforms saving $80K/month — I turn complex data into intelligent, scalable solutions.

ML Engineer • Data Engineer • AI Systems Builder • Computer Vision
View My Work Get In Touch Download Resume
0+
Years Experience
0K+
Monthly Requests
$0K
/mo Savings Delivered
0.9
GPA @ WPI
Scroll
Featured Systems

Production ML Systems

Real-world AI systems I've built and deployed at scale.

LLM Medical Coding System
→ Processes 100K+ monthly requests
→ Used by 10+ enterprise healthcare clients

Semantic Search Platform
→ Indexed 940K healthcare documents
→ Delivered $80K/month operational savings

Document Triage ML Pipeline
→ Processes 900+ documents/day
→ Automated 80% of workflows

About Me

Building Intelligent Systems

From research to production — I bridge the gap between ML models and real-world business value.

sheroz@ml-studio ~ %
>>> import sheroz
>>> sheroz.about()
{
  "name": "Sheroz Shaikh",
  "role": "ML Engineer",
  "education": "M.S. Data Sci @ WPI",
  "gpa": 3.9,
  "focus": [
    "Production LLM Systems",
    "Healthcare AI",
    "Data Engineering",
    "Semantic Search & RAG"
  ],
  "superpower": "Research → Production"
}
>>>

From Models to Business Value

I'm an ML Engineer with 8+ years of experience shipping production AI systems in healthcare and enterprise domains. With an M.S. in Data Science from WPI (3.9 GPA), I specialize in building LLM-powered applications, semantic search platforms, and scalable data pipelines that deliver measurable ROI.

At CrowdANALYTIX, I built a semantic search platform over 940K healthcare documents that saved $80K/month in operational costs. I deploy models that serve 100K+ monthly requests with production-grade monitoring, validation, and error handling.

Production LLM Systems
Healthcare AI / NLP
Semantic Search & RAG
Data Pipeline Design
Time-Series Forecasting
ML Observability
🏆 Best Data Science Project (Winter 2024) — Won 1st place for a healthcare project at WPI, recognized for technical rigor and real-world impact.
Experience

Where I've Shipped AI

Building production ML systems that drive real business outcomes and operational savings.

Biological Information OS, Inc
Machine Learning Engineer
Remote, Florida
Dec 2025 – Present
Python vLLM Neo4j spaCy MongoDB FastAPI Docker Pydantic Airtable API
  • Built a parallelized PDF extraction pipeline for 300+ life-science textbooks that routes each document by content type, vLLM-served vision OCR for scanned/equation-heavy pages, a lightweight parser for native-text PDFs, cutting per-document cost and lifting throughput on self-hosted infrastructure
  • Built a knowledge graph of 1.3K+ topics and 4.1K+ cross-document links from content, exposed via API for downstream search and discovery
  • Built an automated text-analytics pipeline scoring 50+ readability and complexity metrics per page, enabling grade-level and content-difficulty insights across the full corpus
Biogas Engineering
Machine Learning Engineer
Remote, California
May 2025 – Nov 2025
PyTorchFastAPIDockerAWSPrometheusGrafanaClaude/OpenAI API
  • Delivered an automated document-triage service processing 900+ documents daily at 99%+ uptime, replacing manual review across production workflows
  • Automated ticket categorization and prioritization with an LLM-powered routing service, eliminating ~$700/month in manual operational cost
  • Deployed a sensor forecasting pipeline predicting equipment failures up to 30 days ahead, enabling proactive maintenance and reducing unplanned downtime
ProsperOn Graduate Consulting
Data Engineer & Analytics
Remote, Boston
Feb 2025 – Apr 2025
PythonSQLPolarsPandasAWS S3PlotlyPostHog
  • Built a unified analytics warehouse from fragmented event data, cutting reporting latency 75% and standardizing cohort and retention metrics across product teams
  • Shipped self-service dashboards for user adoption and engagement, eliminating recurring ad-hoc analysis requests
Discern Health Graduate Consulting
ML Engineer
Remote, Texas
Aug 2024 – Dec 2024
PythonPySparkAWS SageMakerH2O AutoMLScikit-learnPandasSQL
  • Optimized a production ETL pipeline over 15M+ Medicare records, improving query performance 58%, cutting storage 42%, and reducing repeated scans 75%
  • Improved clinical prediction recall 23% through AutoML-driven feature engineering, surfacing high-impact predictors for patient-outcome models
CrowdANALYTIX
Senior Machine Learning Engineer
Bangalore, India
May 2018 – Jul 2023
PythonLLMsPyTorchApache AirflowFastAPISQLDockerKubernetesPrometheusGrafana
  • Deployed and maintained a production LLM system for automated ICD-10 medical coding, serving 10+ enterprise clients and 100K+ requests/month with built-in validation and monitoring
  • Built a semantic search pipeline across 940K documents, eliminating manual review of 500K+ documents/month and delivering ~$80K/month in operational savings
  • Automated ETL workflows over 2M+ monthly records, reducing manual data operations 70% and saving ~$2.5K/month
  • Led ML systems end-to-end from requirements to production and mentored 3 junior engineers on design and scalable architecture
Motilal Oswal Financial Services
Senior Executive — MIS & Analytics
Mumbai, India
Jul 2017 – Jan 2018
Market Xcel
Research & Data Analyst
Mumbai, India
Aug 2016 – Jul 2017
Tech Stack

My ML Toolkit

The frameworks, tools, and platforms I use to build and ship production AI systems.

AI & Agentic Systems
Large Language Models OpenAI Anthropic Claude Gemini Qwen OpenRouter Ollama LLM Agents LangGraph LangChain LangSmith ReAct Plan-and-Execute RAG Agentic RAG Query Decomposition Tool Calling Structured Output Adaptive Retrieval BM25 Hybrid Retrieval (RRF) Cross-Encoder Reranking Semantic Search Embeddings Vector Databases (FAISS, Qdrant, Pinecone, Chroma) Prompt Engineering Prompt Injection Defense vLLM Hugging Face
ML, NLP & Deep Learning
Machine Learning Deep Learning Natural Language Processing Transformers PyTorch Scikit-learn XGBoost LightGBM H2O AutoML PEFT / LoRA spaCy NLTK fastText Feature Engineering Time-Series Forecasting Anomaly Detection Classification & Regression Reinforcement Learning (DQN, Rainbow) Gymnasium / ALE Computer Vision / OCR Model Training & Evaluation LLM Evaluation (RAGAS, DeepEval, LLM-as-Judge) Weights & Biases Hydra
Data Engineering & ETL
PySpark Polars Pandas NumPy Apache Airflow SQL PostgreSQL MongoDB Redis DuckDB MinIO Pandera Plotly Kafka Data Pipelines ETL Data Warehousing Data Quality Validation Schema Optimization Distributed Computing Workflow Automation
MLOps & Production
FastAPI Celery SQLAlchemy Pydantic Docker Docker Compose Kubernetes Nginx Alembic REST API Design Server-Sent Events (SSE) CI/CD (GitHub Actions) MLflow Prefect DVC Model Deployment & Serving Triton ONNX Ray Batch & Scheduled Pipelines
Cloud & Observability
AWS EC2 AWS S3 AWS Lambda AWS SageMaker GCP OpenTelemetry Prometheus Grafana Loki Tempo LangSmith Pipeline Monitoring Metrics Tracking & Alerting Cloudflare Tunnels
Languages, Tools & Practices
Python SQL JavaScript React Bash Git GitHub Linux uv ruff mypy pytest loguru hatchling PyInstaller Jupyter Unit & Integration Testing Agile
Personal Projects

Research & Open Source

Systems I designed and built end-to-end — from research to deployed, observable platforms.

Single-Page OCR Microservice
A production-ready skeleton for a single-page OCR service: PP-OCRv6 on ONNX Runtime behind a sync FastAPI, with hybrid CPU+GPU workers sharing one Redis queue, content-hash dedup and single-flight caching, and full OpenTelemetry → Prometheus / Loki / Tempo → Grafana observability. It returns a fixed 3-key JSON contract and handled 2,800 concurrent requests with zero errors. Built to be forked as a starting point — bring your own model, or add a frontend as needed.
PP-OCRv6ONNX RuntimeFastAPIRedisDockerGrafana
Agentic RAG Component Ablation Code
A controlled ablation isolating which agentic-RAG components actually drive accuracy on 5,000 HotpotQA multi-hop questions. The full pipeline gains +10.1 EM / +7.6 F1 over baseline, and a fixed-hybrid variant reaches EM 55.0 / F1 63.5 — all on a fully local 7B model at zero API cost, with paired significance tests and bootstrap CIs. Written up as an arXiv paper.
Agentic RAGQwen2.5-7BQdrantDuckDBFastAPI
Production AI Refund Agent
A production e-commerce refund agent built on a deterministic 5-node LangGraph pipeline that enforces a written refund policy via Claude tool forcing at temperature=0. Decisions stream back over SSE, every run is auditable in a JWT-protected trace dashboard, and the system ships with prompt-injection defense, graceful fallback, and full OpenTelemetry → Prometheus / Loki / Tempo → Grafana observability.
LangGraphClaudeFastAPICeleryReact 19
Predictive Maintenance Platform
End-to-end ML platform for turbofan engine remaining-useful-life forecasting, failure classification, and anomaly detection on NASA CMAPSS. LightGBM RUL forecast at RMSE 10.35 (R² 0.939) and a failure classifier at F1 0.918 / AUC 0.996, orchestrated with Prefect, tracked in MLflow, and fully monitored — cloud-deployable on a free-tier VM.
LightGBMPrefectMLflowFastAPIDocker
Retail Demand Allocator
ML platform that forecasts daily retail revenue and optimally allocates marketing budget across channels on the UCI Online Retail dataset. A LightGBM forecaster (benchmarked against H2O AutoML) feeds an evolutionary budget optimizer, with 80 engineered features, a Prefect training pipeline, MLflow + MinIO tracking, and full Prometheus / Grafana monitoring.
LightGBMH2O AutoMLPrefectMLflowDuckDB
Mini-Rainbow DQN Live Demo
End-to-end deep reinforcement learning platform that trains DQN, Double DQN, Dueling DQN, and Rainbow-Lite agents to play Atari Breakout, then runs two agents side-by-side in a live dashboard with SSE frame streaming, Weights & Biases experiment tracking, and Prometheus / Grafana observability.
PyTorchReinforcement LearningGymnasiumFastAPIW&B
Parameter-Efficient Fine-Tuning (LoRA) for Text Classification
Built an end-to-end text classification pipeline with LoRA fine-tuning, using a controlled experimental setup with reproducible preprocessing, training, and benchmarking. Demonstrated efficient adaptation of large language models for domain-specific tasks.
PyTorchLoRA/PEFTTransformersNLP
Open-Source Python Packages (PyPI)
Published 4 production-grade Python packages for ML pipeline profiling, structured logging, and data transformation workflows, enabling teams to standardize monitoring and debugging across ML systems. Available on PyPI for community use.
PythonPyPIML OpsOpen Source
Consulting

Consulting Engagements

Independent engagements building production tools and libraries for clients. Client names are withheld under confidentiality; descriptions are kept high-level, and links are shown only for components released publicly.

PDF Language Detection & Splitting
A production-grade, pure-Python library that detects the language of every page in a PDF (176 languages via fastText), groups consecutive pages into language segments, and splits the document into per-language files entirely in memory — with batch processing and strict type safety.
fastTextpypdfNLPPython
LLM Document Classifier
A Gemini-powered classifier that analyzes a full PDF in a single API call, detects its document type, identifies the relevant pages, and builds a page-to-document-ID map — used as a cost gate in front of a more expensive downstream extraction step.
GeminiPydanticDocument AIPython
Document Pipeline Preflight
An in-memory preflight toolkit for document pipelines: validate PDF integrity, render pages to images, and classify pages as blank vs. content through a 4-signal pure-NumPy cascade — three independent modules shipped as a single pure-Python wheel with 137 tests.
pypdfpypdfium2NumPyPillow
Product Attribute Structuring Tool Private
A batch tool that ingests raw product Excel files, splits them by taxonomy path, applies a tracker-driven column plan sourced from a live spreadsheet (cached in DuckDB and refreshed only on change), and outputs clean files alongside a full audit report — parallelized, with a headless API.
PythonDuckDBExcelAutomation
SKU Batch Summary Generator Private
A desktop tool that computes monthly unique and non-unique SKU counts from multiple batch CSV sources, with approved-batch filtering and cross-year date ranges. Distributed as a Windows executable built and released automatically through a CI pipeline.
PythonPandasDesktop AppCI/CD
Taxonomy Processing Pipeline Private
An automated pipeline that consolidates many Excel/CSV files into a structured taxonomy — extracting category and attribute columns and units of measure, deduplicating, and merging with an existing taxonomy. Packaged as a Windows executable for non-technical operations teams.
PythonPandasDesktop AppAutomation
Education

Academic Foundation

M.S. Data Science
Worcester Polytechnic Institute (WPI)
Worcester, MA May 2025 GPA: 3.9 / 4.0
Best Data Science Project (Winter 2024) — 1st place for healthcare project
B.E. Electronics & Telecommunication
AIKTC School of Engineering and Technology
Mumbai, India May 2016
Contact

Let's Build Intelligent Systems Together

Whether it's deploying LLMs at scale, building data pipelines, or discussing the future of AI in healthcare — let's connect.

shaikh.sheroz07@gmail.com +1 (781) 655-9092 sherozshaikh sherozshaikh PyPI
Send Me a Message