Machine Learning & AI
Models that survive contact with production.
GANs, diffusion models and LLM systems in PyTorch and TensorFlow, taken past the notebook into deployed inference on AWS SageMaker and GCP Vertex AI.
- Generative
- WGAN-GP · Conditional DDPM · PaLM-E
- Frameworks
- PyTorch · TensorFlow · Keras
- Serving
- SageMaker · Vertex AI · RunPod
Data Engineering
Pipelines built to be boring.
End-to-end ETL orchestrated in Apache Airflow, with automated quality gates and parallel execution across roughly 10 GB of records.
- Orchestration
- Apache Airflow · BullMQ
- Stores
- PostgreSQL · DynamoDB · S3 · DuckDB
- Scale
- ~10 GB EHR · 3× throughput
Full-Stack Development
The surface that makes it usable.
Typed Next.js frontends over REST APIs in ASP.NET MVC and Node, with role-based access control where the data needs it.
- Frontend
- Next.js · TypeScript · React
- Backend
- ASP.NET MVC · C# · Fastify
- Contracts
- REST · role-based access control
IoT & Embedded
Where the data actually comes from.
Arduino and Raspberry Pi sensor firmware, local aggregation and real-time dashboards for the hardware that produces the data in the first place.
- Hardware
- Arduino · Raspberry Pi
- Telemetry
- Real-time aggregation · anomaly alerts
- Shipped
- SHEMS energy monitor
Aditya PatelPractice
From researchto production.
Most of the hard problems live outside the model file: data you cannot trust, a pipeline that has to run at 4am, and a deployment nobody wants to be paged about.
4
Industry internships
AI · Data Eng · Web · IoT
9
Projects shipped
ML, ETL, IoT & web
50K+
Synthetic records
SynMedix on SageMaker
84
Instruments modelled
Global Market Lab
Aditya PatelCapabilities
Technical stack.
Everything listed here has shipped in something real: an internship deliverable, a deployed model, or a project running today.
Languages
- Primary
- Python
- Backend
- C# · C / C++
- Web
- TypeScript · JavaScript
- Query
- SQL
ML / AI
- Frameworks
- PyTorch · TensorFlow · Keras
- Generative
- GANs · Diffusion · LLMs
- Classical
- Scikit-learn
- Multimodal
- PaLM-E · cross-attention
Data & Cloud
- Orchestration
- Apache Airflow
- Relational
- PostgreSQL · SQL Server
- NoSQL
- DynamoDB · Redis
- Object store
- AWS S3
Web & Systems
- Frontend
- Next.js · React
- Backend
- ASP.NET MVC · Fastify · Node
- Interfaces
- REST APIs
- Embedded
- Arduino · Raspberry Pi
Tooling
- Containers
- Docker
- CI / CD
- GitHub Actions · Git
- Analysis
- Pandas · NumPy
- Reporting
- Power BI
Deployments
- AWS
- SageMaker · S3 · DynamoDB
- GCP
- Vertex AI
- GPU
- RunPod RTX 4090
- Edge
- Vercel
Featured BuildSynMedix AI
Synthetic patient data, at scale.
A distributed EHR processing platform and the generative model on top of it. Three steps from raw records to a dataset a research team can train on.
- 01The problem
Ten gigabytes of records nobody could touch
Large volume, absolute privacy constraints, and serial processing slow enough that iterating on a model means waiting overnight.
- Input
- ~10 GB electronic health records
- Constraint
- No real patient data downstream
- 02The pipeline
Parallel execution, then a generative layer on top
Ingest restructured around parallel execution, cutting turnaround to a third. A generative layer on the cleaned corpus learns the joint distribution rather than copying any single record.
- Throughput
- 3× over the serial baseline
- Stack
- Python · SQL · PyTorch · TensorFlow
- 03The outcome
50,000+ synthetic records, zero real patients exposed
Deployed on SageMaker, emitting records that keep the statistical structure of the source corpus without carrying any individual through it.
- Generated
- 50,000+ synthetic patient records
- Deployment
- AWS SageMaker
Aditya PatelContact
Let's buildsomething.
A hard ML problem, a pipeline that won't scale, or a role where you need someone who will actually dig in. Leave an address and I'll reply within 24 hours.
Prefer the long form? Full contact page