About Projects Experience Contact
Available for Opportunities, 2026 · Boston · Open to Relocation

Hrishikesh Keswani.

Machine Learning Engineer & Data Scientist

MS in Data Analytics Engineering from Northeastern. I build scalable ML systems — from RAG pipelines and LLM inference optimization to distributed training frameworks and event-driven cloud architectures. I ship production-grade AI that actually works.

View My Work
Stack
PyTorch + AWS
Education
MS · Northeastern . Dec 2025
Available
Full-time
PyTorch ✦ LangChain ✦ vLLM ✦ AWS ✦ Kubernetes ✦ FastAPI ✦ FAISS ✦ PostgreSQL ✦ Docker ✦ Ray ✦ HuggingFace ✦ DeepSpeed ✦ PyTorch ✦ LangChain ✦ vLLM ✦ AWS ✦ Kubernetes ✦ FastAPI ✦ FAISS ✦ PostgreSQL ✦ Docker ✦ Ray ✦ HuggingFace ✦ DeepSpeed ✦

// About me

Turning data
into intelligent
systems.

I'm an ML engineer who thrives at the intersection of research and production. I've built RAG pipelines, optimized LLM inference at scale, and shipped distributed training frameworks. I care about clean abstractions, reproducible experiments, and systems that hold up under real load.

~/hrishikesh - zsh
› cat profile.json
{
  "name": "Hrishikesh Keswani",
  "role": "Machine Learning Engineer",
  "education": "MS Data Analytics Eng, Northeastern",
  "stack": ["PyTorch", "LangChain", "AWS", "vLLM"],
  "open_to": "FAANG & great AI startups",
  "status": "building...",
}
›

// Experience

Where I've worked.

Jan 2025 – July 2025
Machine Learning Engineer
Boehringer Ingelheim · Athens, GA
Built a scalable RAG system with Llama-3, GPT-4o, and Nova Lite to analyze 10K+ surveys, driving leadership insights. Increased LLM inference throughput from 1.2K to 2.0K tokens/sec using vLLM batching and multi-GPU Ray. Deployed monitoring with Grafana and Prometheus, optimizing p95 latency and GPU utilization.
Pharma
Aug 2022 – July 2023
Data Scientist
Yunometa · Mumbai, India
Developed a multilingual document text detection model using Faster R-CNN (ResNet-50) in PyTorch, improving extraction recall by 30%. Architected an event-driven AWS pipeline (S3, Lambda, SageMaker) for PDF ingestion and scalable inference.
ML
Dec 2021 – Mar 2022
Data Scientist
Kritexco · Mumbai, India
Analyzed 50K+ NFT marketplace transactions to detect fraud. Built fraud detection models using Logistic Regression and Random Forest, achieving PR-AUC of 0.82 and recall of 0.78 on highly imbalanced data using SMOTE.
Fintech
June 2021 – Aug 2021
Machine Learning Engineer
Aadyam Infotech · Pune, India
Engineered an NLP pipeline leveraging Azure AI and regex to parse and extract key information from unstructured documents, achieving 86% accuracy on extracted records.
NLP

// Projects

Things I've built.

FastAPI · pgvector · Groq · React
AI Analyst Copilot
Enterprise Business Intelligence Agent
Built an end-to-end NL→SQL pipeline that converts natural language questions into SQL queries, executes them against PostgreSQL, and generates executive-level insights. Implemented RAG-based schema linking from scratch using pgvector and sentence-transformers — no LangChain. Achieved 80% execution accuracy on a 20-query benchmark. Includes a feedback loop for SQL correction and a React dashboard served via FastAPI.
RAG
Python · LangChain · FAISS · GCP
LinkedIn Search Optimizer
RAG-based Search Engine
Built a RAG-based search engine analyzing 124K+ LinkedIn posts to deliver contextual job search via natural language queries. Deployed with Docker, Kubernetes, and FastAPI with ELK monitoring, achieving 0.8 Recall@10. Distributed Airflow pipeline on GCP with TTL-based vector refresh to sustain search relevance across multi-source data.
RAG
LangChain · FAISS · Streamlit
Resume RAG Chatbot
AI-powered Resume Q&A
Built a conversational RAG chatbot that lets anyone ask questions about my experience, skills, and projects in natural language. Powered by LangChain, FAISS vector search, and an LLM backend — deployed live on Streamlit Cloud.
RAG
DeepSpeed · Ray · Docker · HuggingFace
Distributed ML Training Framework
Multi-GPU LLM Training
Engineered a scalable multi-GPU training framework leveraging DeepSpeed ZeRO optimization and Ray orchestration for memory-efficient parallel LLM training. Built a modular pipeline with automated hyperparameter tuning and containerized deployment supporting domain-specific adaptation at scale.
LLM

// Contact

Let's build
something great.

I'm always open to interesting conversations, new opportunities, and collaborations. If you have a project in mind or just want to say hello, email me below.

keswani.hrishikesh7@gmail.com
Hrishikesh_Keswani_Resume.pdf
↗ Open ↓ Download