Skip to content
View krishnau225's full-sized avatar
  • United States
  • 05:44 (UTC -04:00)

Block or report krishnau225

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
krishnau225/README.md

Typing SVG


AI and software engineering



Portfolio Resume


LinkedIn Email GitHub


Experience Focus Location Profile Views


About Me

I am an AI/ML Engineer with 5+ years of experience designing, training, optimizing, and deploying large-scale artificial intelligence systems.

My work spans foundation models, multimodal LLMs, distributed model training, Generative AI, retrieval-augmented generation, RLHF, DPO, AI agents, MLOps, and scalable GPU inference.

I build end-to-end AI platforms covering:

  • Large-scale model training and evaluation
  • Distributed GPU infrastructure
  • Retrieval-augmented generation
  • Preference optimization and model alignment
  • Agentic AI and multi-agent workflows
  • Model serving and inference optimization
  • Data engineering and feature pipelines
  • Cloud-native MLOps and observability

I focus on combining advanced AI research with reliable software engineering to create secure, scalable, measurable, and production-ready systems.

Current interests: multimodal foundation models, agentic AI, distributed inference, model efficiency, GPU optimization, and enterprise Generative AI.


Technical Expertise

Skills

View complete technical stack
Area Technologies
Programming Python, SQL, Bash, Java, C++, CUDA
AI and Machine Learning PyTorch, TensorFlow, Scikit-learn, XGBoost, OpenCV, YOLO
Generative AI LLMs, Multimodal AI, Foundation Models, MoE, SFT, RLHF, DPO, LoRA, PEFT
AI Agents LangChain, LangGraph, LlamaIndex, MCP, Multi-Agent Systems, Tool Calling, Function Calling
RAG and Retrieval RAG, Embedding Models, Vector Search, FAISS, Prompt Engineering
Distributed Training PyTorch Distributed, FSDP, DeepSpeed, Megatron-LM, Ray, NCCL
Model Serving vLLM, TensorRT-LLM, Quantization, Tensor Parallelism, Pipeline Parallelism
Data Engineering Apache Spark, Databricks, Airflow, Kafka, ETL, Streaming Pipelines, Synthetic Data
Cloud Platforms AWS, Microsoft Azure, Google Cloud Platform
Infrastructure Docker, Kubernetes, AWS EKS, REST APIs, Microservices, Service Mesh, Terraform
MLOps MLflow, GitHub Actions, ArgoCD, CI/CD, Model Monitoring, Model Versioning
Monitoring Prometheus, Grafana, LangSmith, Experiment Tracking
Databases and Analytics PostgreSQL, MySQL, SQL, Power BI
Evaluation MMLU, GPQA, HumanEval, A/B Testing, Custom Model Benchmarking, XAI

Professional Impact

🧠 Foundation Models ⚡ AI Infrastructure 🔍 Enterprise AI 🏭 Predictive ML
18% benchmark improvement 32% faster training 24% better retrieval accuracy 32% less downtime
21% higher human evaluation 35% lower inference cost 600M+ daily users supported 38% lower latency
Trillion-token pipelines Distributed GPU platforms RAG and agentic workflows 15,000+ industrial assets

What I Have Built

🧠 Foundation Models
   Built multimodal model-training pipelines processing trillions of tokens
   across distributed GPU infrastructure.

⚙️ Mixture-of-Experts
   Optimized expert routing and distributed training using PyTorch, CUDA,
   FSDP, and Megatron-LM.

🎯 Model Alignment
   Implemented RLHF, DPO, and preference-optimization workflows for safer,
   more relevant production AI systems.

🔍 Enterprise RAG
   Developed embedding, FAISS, vector-search, and transformer-based retrieval
   systems for enterprise knowledge workloads.

🤖 Agentic AI
   Built scalable AI-agent platforms with RAG, workflow orchestration,
   tool calling, conditional execution, and distributed services.

🚀 Model Serving
   Deployed large-scale inference systems using vLLM, TensorRT-LLM,
   Kubernetes, autoscaling, and GPU optimization.

📊 MLOps
   Implemented automated training, deployment, monitoring, evaluation,
   model versioning, and governance pipelines.

🏭 Industrial Machine Learning
   Built predictive-maintenance, computer-vision, IoT-streaming, and
   digital-twin solutions for industrial environments.

Selected AI Projects

Helix — Production-Grade AI Inference Gateway

Language Last Commit

A production-grade LLM gateway that provides one secure interface for OpenAI, Ollama, and Anthropic models. It combines score-based provider routing, circuit-breaker fallback, pgvector semantic caching, per-tenant JWT authentication, Redis rate limiting, SSE token streaming, structured request logging, and Prometheus observability.

The gateway was deployed on Fly.io and validated through a cache-warmed k6 benchmark with 3,885 requests, zero errors, 100 concurrent virtual users, 18.43 requests per second, and 567 ms p95 latency. A React dashboard provides live visibility into latency, token usage, provider health, cache performance, and estimated inference cost.

Go React TypeScript OpenAI Ollama Anthropic PostgreSQL pgvector Redis Docker Prometheus k6


Multi-Agent Code Repair — Autonomous Issue-to-Patch Engineering System

Language Last Commit

A stateful multi-agent software-repair system that converts a GitHub issue into a validated code patch. Specialized agents research the repository, identify relevant files, plan a minimal change, generate structured search-and-replace edits, and validate the result inside an isolated Docker environment.

The workflow pins every run to a specific commit, grounds patches in verbatim source code, executes pytest, Ruff, and Mypy checks, retries using structured validation feedback, persists progress through SQLite checkpointing, and can optionally open a GitHub pull request. A real-world Click issue run completed with 1,436 tests passing, clean linting and type checks, zero retries, and approximately $0.35 in model cost.

Python LangGraph LiteLLM Docker PyGithub Pydantic pytest Ruff Mypy SQLite


ResearchFlow AI — Self-Correcting Agentic Research Platform

Language Last Commit

A production-oriented agentic research system that decomposes complex questions, retrieves web evidence, evaluates evidence quality, and synthesizes citation-grounded answers.

Its LangGraph workflow coordinates Planner, Decomposer, Search, Verifier, and Synthesizer nodes. When retrieved evidence falls below a confidence threshold, the verifier routes execution back to search with a refined query instead of generating an answer from weak context. The application exposes the workflow through FastAPI and provides a React interface with citations and a visible execution trace.

Python LangGraph FastAPI Tavily Ollama React TypeScript Agentic AI


ECG-PEFT Bench — Parameter-Efficient Medical AI Benchmark

Stars Last Commit

A reproducible medical AI benchmarking framework that compares LoRA and frozen-encoder adapter fine-tuning across Wav2Vec2, HuBERT, and ECG-FM backbones for binary ECG-segment classification.

The pipeline applies model-aware ECG preprocessing, backbone-specific resampling, mask-aware mean pooling, consistent data splits, controlled random seeds, automated evaluation, and exact confusion-matrix reporting. By holding training conditions constant across six configurations, the benchmark isolates performance differences caused by the backbone and adaptation strategy rather than experimental variation.

Python PyTorch Hugging Face PEFT LoRA Wav2Vec2 HuBERT ECG-FM


DocuMind — Explainable RAG Document Intelligence

Language Last Commit

A production-oriented explainable RAG platform that allows users to upload PDFs and receive answers grounded strictly in retrieved document content.

DocuMind extracts and cleans PDF text, creates configurable overlapping chunks, generates OpenAI embeddings, and stores document-scoped vectors in Pinecone. Every query is filtered by document ID to prevent cross-document retrieval, and responses include the supporting chunks for inspection. The system explicitly handles empty retrieval results and uses LlamaIndex relevancy evaluation to assess answer quality without requiring a labeled dataset.

Python FastAPI Pinecone OpenAI LlamaIndex RAG React TypeScript


ContextFlow AI — Privacy-First Local Browser Agent

Language Last Commit

A privacy-first Chrome extension that brings local LLM intelligence directly into the browser without cloud inference or external API keys.

The extension extracts live webpage context for grounded question answering, summarizes complete pages, and explains highlighted text through a right-click context-menu action. A Manifest V3 background service worker coordinates requests between the webpage, React popup, injected floating panel, and a locally running Ollama model, ensuring that browsing content remains on the user’s machine.

TypeScript React Chrome Manifest V3 Ollama Llama 3.2 Vite Local AI


Additional Projects

A workforce-management platform designed to coordinate employee schedules and duty assignments, including concurrency-aware updates for safer shift management.

A parameter-efficient conditional diffusion research project focused on reducing trainable parameters and improving training efficiency while preserving generation quality.

A text-to-audio generation pipeline that uses CLAP-based semantic scoring and multi-candidate selection to improve alignment between generated audio and input prompts.


Professional Experience

Organization Position Period Focus
Meta AI/ML Engineer February 2025 – Present Foundation models, multimodal AI, distributed training, RLHF, RAG, MLOps and large-scale inference
Accenture Machine Learning Engineer March 2021 – June 2024 Predictive maintenance, computer vision, IoT analytics, data engineering, cloud AI and MLOps

Experience Highlights

Meta — AI/ML Engineer

  • Built multimodal foundation-model training pipelines using Python, PyTorch, Transformers, and distributed GPU infrastructure.
  • Developed Mixture-of-Experts architectures using CUDA, FSDP, and Megatron-LM.
  • Implemented RLHF, DPO, and preference-optimization workflows.
  • Designed petabyte-scale data pipelines using Spark, Ray, Airflow, and Python.
  • Built RAG systems using FAISS, embeddings, vector search, and transformer models.
  • Developed model-evaluation frameworks using MMLU, GPQA, HumanEval, and custom benchmarks.
  • Deployed containerized inference platforms serving large-scale global workloads.
  • Optimized inference using vLLM, TensorRT-LLM, autoscaling, and GPU-serving strategies.
  • Implemented MLOps and CI/CD using MLflow, GitHub Actions, ArgoCD, Prometheus, and Grafana.
  • Built agentic AI systems integrating RAG, workflow orchestration, tool calling, and distributed services.

Accenture — Machine Learning Engineer

  • Built predictive-maintenance systems using real-time IoT sensor data.
  • Developed TensorFlow and XGBoost models for industrial failure prediction.
  • Optimized model latency, cloud inference cost, and feature-engineering workflows.
  • Designed Spark, Databricks, SQL, and Airflow data pipelines.
  • Developed computer-vision inspection systems using OpenCV, PyTorch, and YOLO.
  • Implemented end-to-end MLOps pipelines with MLflow, Docker, Kubernetes, and CI/CD.
  • Built cloud-native AI platforms using AWS SageMaker, S3, Glue, Lambda, and EKS.
  • Developed Java, Spring Boot, and FastAPI prediction microservices.
  • Created digital-twin analytics using AWS IoT Core and streaming telemetry.
  • Delivered Power BI dashboards for machine health, anomaly detection, and predictive maintenance.

Education

Degree Institution
Master of Science in Computer Science and Engineering University of South Florida

Certifications

  • AWS Certified Machine Learning Engineer – Associate
  • AWS Certified AI Practitioner
  • Microsoft Certified: Azure AI Engineer Associate
  • Databricks Certified Machine Learning Professional

GitHub Activity


GitHub statistics Most-used languages
GitHub contribution snake

Areas of Interest

Foundation Models             Multimodal Artificial Intelligence
Generative AI                 Agentic AI and Multi-Agent Systems
Distributed Model Training    Retrieval-Augmented Generation
Model Alignment               GPU Inference Optimization
MLOps and Model Governance    Cloud-Native AI Infrastructure
Computer Vision               AI Evaluation and Benchmarking

Pinned Loading

  1. krishnau225 krishnau225 Public

  2. multi-agent-code-repair multi-agent-code-repair Public

    Autonomous GitHub issue → validated patch pipeline: LangGraph agents research, plan, and generate search-replace patches, validated in a Docker sandbox (pytest · ruff · mypy), with structured retry…

    Python

  3. helix helix Public

    A production-grade AI inference gateway built in Go that routes requests across multiple LLM providers with semantic caching, JWT authentication, rate limiting, circuit-breaker fallback, SSE stream…

    Go

  4. DocuMind DocuMind Public

    Docu-Mind is an AI-powered document assistant that answers questions from PDFs using natural language. It combines RAG, GPT-4, and Pinecone to deliver fast, accurate, and verifiable responses throu…

    TypeScript

  5. langgraph-research-agent langgraph-research-agent Public

    Self-correcting AI research agent built with LangGraph, Tavily Search, Ollama, FastAPI, and React. Decomposes complex questions, retrieves web evidence, verifies relevance, and generates citation-g…

    Python

  6. ecg-peft-benchmark ecg-peft-benchmark Public

    Benchmarking PEFT (LoRA vs adapters) for ECG segment classification using Wav2Vec2, HuBERT, and ECG-FM with full metric reporting and ablations.

    Jupyter Notebook 1