Open to new opportunities

Priyam Kakati

AI scientist at XO Health. I build multi-agent systems, retrieval-augmented generation pipelines and computer-vision models, and take them all the way to production.

View résumé Email me
  • Top 1%Kaggle Expert, worldwide
  • 4+years building ML systems
  • 60k+learners mentored
Nearest to the query pointk = 5

    My projects and tools, clustered by topic like an embedding space. Move across the map to query it, then select a result to jump to it.

    Selected work

    Ten projects across agentic AI, generative AI, computer vision and classical machine learning.

    • Agents Project lead

      MCP autonomous recruiter agents

      A multi-agent system built with AutoGen and the Model Context Protocol. The agents work through resume parsing, candidate enrichment and recruiter Q&A on their own.

      • AutoGen
      • MCP
      • GPT-4
      • LangChain
      25%lower recruitment MTTR
    • LLMs & NLP

      Enterprise RAG pipeline

      A scalable retrieval-augmented generation system using Pinecone and Weaviate for vector storage, with advanced chunking and BGE embeddings.

      • Pinecone
      • Weaviate
      • LlamaIndex
      • BGE embeddings
      +35%document Q&A accuracy
    • LLMs & NLP

      Custom instruction fine-tuning with QLoRA

      Fine-tuned Llama 3 8B for an industry compliance use case using QLoRA (PEFT) on custom datasets.

      • Llama 3
      • QLoRA / PEFT
      • PyTorch
      • Custom datasets
      40%fewer hallucinations on compliance queries
    • LLMs & NLP

      Intelligent document extraction

      High-accuracy extraction for invoices and claims, combining OCR with large language models (GPT-4 and Claude).

      • OCR
      • GPT-4
      • Claude
      • AWS Lambda
      80%operational efficiency gain
    • Agents

      Financial research agent swarm

      A team of specialized agents that collaborate to write detailed, cited market research reports automatically, using dynamic tool calling.

      • LangChain agents
      • Claude 3
      • Tool use
      • Autonomous workflows
      Agent roles
      1. Planner
      2. Researcher
      3. Editor
      4. Critic
    • Vision

      Large-scale object detection

      Fine-tuned YOLOv5 and Mask R-CNN models deployed for industrial defect detection, running on EC2.

      • YOLOv5
      • Mask R-CNN
      • AWS EC2
      • MLOps
      1,000+parallel inference jobs
    • Classical ML

      HFT time-series predictor

      Time-series forecasting with XGBoost and LightGBM to predict short-term price moves in a simulated high-frequency trading environment, optimized for latency.

      • XGBoost
      • LightGBM
      • Time series
      • Pandas / NumPy
    • LLMs & NLP

      Zero-shot text classification

      An API that classifies text against labels it has never seen, using pre-trained Hugging Face Transformers (BART and NLI models) with no retraining.

      • Transformers
      • BART / NLI
      • NLP
    • Vision

      Real-time lane detection

      A fast deep convolutional network for real-time lane detection in self-driving car simulations, tuned for inference speed.

      • PyTorch
      • OpenCV
      • Latency optimization
    • Classical ML

      Customer churn prediction API

      A complete churn prediction pipeline with a scikit-learn ensemble, served as a low-latency Flask microservice with CI/CD on Kubernetes.

      • scikit-learn
      • Flask
      • Kubernetes
      • CI/CD

    Experience

    Four remote roles since 2021, moving from computer vision to conversational AI, retrieval and autonomous agents.

    1. – Present

      AI Scientist Current

      XO Health Inc., remote

      Leading AI-powered data extraction and automation systems built on GPT-4 and Claude. Built scalable ETL workflows that save hundreds of manual hours a month.

    2. –

      Data Scientist

      Korn Ferry, remote

      Architected RAG pipelines on Pinecone and GPT-4o, built MCP-integrated autonomous agents for recruiter workflows and set up ML observability stacks.

    3. –

      Data Scientist

      Gravitas.ai, remote

      Developed conversational AI agents with Lex and LangChain for 10k+ daily users. Set up CI/CD pipelines that cut release cycles by 50%.

    4. –

      Data Scientist

      NEXTGEN Innovation Labs, remote

      Trained YOLOv5 and Mask R-CNN models reaching 92% mAP for computer-vision products. Introduced MLOps practices for scalable deployment and integrated models with MongoDB and BigQuery.

    Toolkit

    What I reach for, from research notebooks to production clusters. Bold entries are where I go deepest.

    Generative AI and LLMs

    • Multi-agent systems (AutoGen)
    • RAG architectures
    • LangChain / LlamaIndex
    • Fine-tuning (PEFT, QLoRA)
    • Hugging Face Transformers
    • GPT and Claude APIs
    • T5 / Whisper / CLIP

    Core ML and frameworks

    • PyTorch / TensorFlow / Keras
    • JAX
    • scikit-learn
    • XGBoost / LightGBM / CatBoost
    • Time-series analysis
    • Statistical modeling

    Computer vision and NLP

    • OpenCV / MediaPipe
    • YOLOv8 / Detectron2
    • Image segmentation and OCR
    • BERT and Transformer models
    • spaCy / NLTK / Gensim
    • Text and intent classification

    Cloud and MLOps

    • AWS (SageMaker, EC2, Lambda, S3, Glue)
    • Azure ML / Data Factory
    • GCP (Vertex AI, BigQuery)
    • Kubernetes (CKAD) and Docker
    • Terraform / GitOps
    • MLflow / Airflow / CI/CD
    • Grafana / Prometheus / CloudWatch

    Data engineering and stores

    • SQL (advanced)
    • Apache Spark / Databricks
    • Apache Kafka
    • MongoDB / PostgreSQL / Redis
    • Pinecone / Weaviate
    • Data preprocessing and ETL

    Languages and tools

    • Python (expert)
    • C++ / Rust / Go
    • Java / Scala / R / MATLAB
    • TypeScript / JavaScript
    • Flask / FastAPI / gRPC
    • Git / GitHub / GitLab

    Certifications

    AWS Certified Machine Learning – Specialty TensorFlow Developer Certificate

    Have a hard ML problem? Let’s talk.

    I’m open to new opportunities in agentic AI, retrieval and computer vision. Email is the quickest way to reach me.