Skip to content
View prathyyyyy's full-sized avatar

Block or report prathyyyyy

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
prathyyyyy/readme.md

About Me

I am a Data Systems and ML/LLM Engineer focused on building reliable systems at the intersection of distributed data engineering, machine learning infrastructure, and applied AI.

My work spans high-throughput batch and streaming pipelines, lakehouse platforms, GPU-accelerated ML/LLM workflows, semantic retrieval, multimodal modeling, deployment, monitoring, and production optimization on AWS and Azure.

I care about measurable performance, reproducible experimentation, robust system design, and taking products from raw data to deployed intelligence.

Open to Opportunities

Data Engineer · ML Engineer · ML Platform Engineer · Applied ML/LLM Engineer · Data Scientist


Engineering Impact

67M+ events processed 4B parameter multimodal model 2M+ vectors indexed
520 equities 33 year backtest 0.81 AUC

Outcome Engineering System
50% faster queries and 40% less storage Apache Hudi optimization for large-scale analytical workloads
0.542 seconds per evaluation sample Sharded multi-GPU evaluation for a 4B-parameter multimodal model
1.4B of 4B parameters active per route Top-2 Mixture-of-Experts routing in KAIROS
0.81 AUC and 0.73 recall Distributed truck-delay prediction with GPU-accelerated XGBoost
IC 0.09 and ICIR 0.62 Regime-aware alpha research across 520 equities

Featured Engineering Work

KAIROS

Multimodal driving-scene reasoning with a hybrid SSM architecture

PyTorch Mamba-2 CfC LoRA DeepSpeed ZeRO-3 MoE

  • Designed and trained a 4B-parameter, linear-complexity multimodal model across 100K+ driving logs.
  • Combined looped Mamba-2 blocks, attention residuals, CfC fusion, and LoRA fine-tuning.
  • Scaled effective depth 4x through weight sharing and top-2 MoE routing.
  • Built a sharded multi-GPU evaluation pipeline reaching 0.542 seconds/sample.

Multi-Signal Alpha Research

Distributed, regime-aware financial ML and backtesting platform

Apache Spark Delta Lake Databricks GBDT Transformers HMM

  • Built a Spark-based Delta Lakehouse spanning feature engineering, modeling, and walk-forward validation.
  • Combined GBDT, Transformer, and Hidden Markov Model signals using 100+ engineered features.
  • Evaluated 520 equities across a 33-year backtest.
  • Achieved IC 0.09 and ICIR 0.62 with regime-aware ensembling.

Streaming Analytics & Security

High-throughput e-commerce intelligence and real-time threat detection

Spark Apache Hudi Kinesis Flink DynamoDB AWS

  • Engineered batch and streaming pipelines processing 67M+ events.
  • Delivered traffic, conversion, cart-abandonment, category, and brand analytics.
  • Improved query performance by 50% while reducing storage by 40%.
  • Built stateful DDoS and bot-detection pipelines with DynamoDB-backed state.

GPU-Accelerated ML Platform

Scalable truck-delay prediction, deployment, and monitoring

Spark ML XGBoost4J-Spark RAPIDS SageMaker Evidently AI

  • Developed a distributed training pipeline with Spark ML and GPU-accelerated XGBoost.
  • Reached 0.81 AUC and 0.73 recall on truck-delay prediction.
  • Productionized orchestration and deployment with SageMaker Pipelines.
  • Added CI/CD, drift monitoring, automated alerting, and evaluation workflows.

Semantic Search & Relevance Platform

Sentence-BERT FAISS Embeddings Vector Search Flask Docker

  • Built a semantic retrieval platform indexing 2M+ vectors with Sentence-BERT and FAISS.
  • Developed scalable embedding, batching, indexing, evaluation, and low-latency top-K retrieval workflows.
  • Exposed retrieval through a containerized API for high-throughput query processing.

Technical Toolkit

Languages, Cloud & Infrastructure

Languages cloud and infrastructure icons

Data, ML & AI

Data machine learning and AI icons
Apache Spark Databricks Delta Lake Apache Flink Apache Hudi Hugging Face FAISS DeepSpeed NVIDIA RAPIDS Amazon SageMaker

Domain Technologies
Data Engineering Apache Spark, PySpark, Kafka, Flink, Kinesis, Databricks, Delta Lake, Apache Hudi, ETL/ELT
ML & LLM Systems PyTorch, Hugging Face Transformers, Spark ML, XGBoost4J-Spark, RAPIDS, LoRA, DeepSpeed, Mamba-2, MoE
Search & Retrieval Sentence-BERT, FAISS, embeddings, vector indexing, semantic search, top-K retrieval
Cloud & MLOps AWS, Azure, SageMaker, Azure ML, Azure Data Factory, Azure DevOps, Evidently AI, Docker, CI/CD
Backend & Storage C#, .NET, Python, SQL, PostgreSQL, Flask, event-driven services

Professional Experience

Software Engineer & AI Engineer Trainee — Thalir Tech

December 2022 – December 2023

  • Implemented backend services with C#/.NET and PostgreSQL for event-driven workflows, scoring systems, and real-time leaderboards.
  • Contributed across software engineering, data pipelines, and ML workflows for production systems deployed on Azure.
  • Collaborated across engineering functions to improve system integration, reliability, and delivery.

Education & Recognition

Education

B.E. in Computer Science and Engineering
Anna University, Tamil Nadu, India
CGPA: 8.1/10

Certifications & Award


What I Want to Build Next

I am interested in teams developing distributed data platforms, real-time analytics, ML/LLM infrastructure, multimodal systems, and large-scale retrieval platforms—especially roles where I can contribute across ingestion, processing, training, evaluation, deployment, observability, and continuous improvement.

Let’s Build Reliable Systems at Scale

Email Prathy



Footer banner

Pinned Loading

  1. Forest-Fire-Detection Forest-Fire-Detection Public

    Forest Fire Detection By Convolutional Neural Network

    Jupyter Notebook 17 4

  2. Medical-Data-Extraction Medical-Data-Extraction Public

    Medical Data Extraction By Pytesseract (Google Optical Character Recognition Engine) and Computer Vision

    Jupyter Notebook 20 3