Open to new roles
  • New York

Hi there, nice to meet you.

Rishi Chhabra

I'm an AI engineer in New York. I build LLM platforms and RAG systems, and the checks that tell you whether they actually work.

Download resume ↓ Email me ↗

01 About me

Portrait of Rishi Chhabra
New York, NY

Right now I'm at Ikka Labs, building the internal AI tools other engineers lean on. The biggest one is Apex, a self-service platform that 4+ teams use to spin up 13 different data and inference engines on their own.

I also built PRaaS, a code-review agent teams can switch on for their repos. These days it reads about 117 pull requests a week.

Before Ikka, I spent my last two semesters at AriesView turning keyword search over 3,000 documents into a proper RAG system, and context precision went from 0.61 to 0.94. I finished my master's in Machine Learning at Stevens in May 2026. My first job was in Delhi, building mobile apps for a startup.

Currently
AI Engineer at Ikka Labs
Based in
New York
Studied
M.S. in Machine Learning, Stevens
Looking for
AI engineering, AI platform, MLOps, or forward deployed roles

Highlights

4+

teams use Apex, the platform I built

Ikka Labs
117

pull requests my review agent reads each week

Ikka Labs
0.61 → 0.94

context precision after I rebuilt retrieval

AriesView
34%

faster API responses after I swapped Redis for Valkey

AriesView
02 Work

Where I've shipped.

New York, NY
  • I built Apex, a self-service platform that 4+ teams use to provision 13 data and inference engines across on-prem Kubernetes, Amazon EKS, and Azure AKS.
  • I made those deployments safe to repeat, with policy checks, signed previews before anything goes out, secrets kept in HashiCorp Vault, and an AI-guided setup flow.
  • I created PRaaS, a code-review agent that runs in GitHub Actions. Teams opt in, and it now reviews about 117 pull requests a week across 13 repositories, picking between seven Gemini, Amazon Bedrock, and self-hosted models.
Ikka Labs · 2026
Boston, MA (Remote)
  • I moved a 3,000-document enterprise corpus off keyword-only search and onto production RAG, with semantic chunking, hybrid BM25 and semantic retrieval in Weaviate, GraphRAG on Neo4j, and Cohere reranking. On a 100-query test set, context precision went from 0.61 to 0.94.
  • I set up the evaluation pipeline with RAGAS, DeepEval, custom metrics, and LLM-as-a-judge, and put it in CI so quality regressions show up in a pull request instead of in production.
  • I shipped a B2B chatbot on top of that retrieval stack and worked directly with the business clients until it was delivered.
  • I swapped Redis for Valkey on our self-hosted caching, and the API endpoints got 34% faster.
AriesView · 2025
Delhi, India
  • I built three cross-platform apps with Flutter, Node.js, MongoDB, and S3: iManager for football teams, Panda for dating, and Paradisseo for events and meeting people. All three launched as MVPs on the App Store and Google Play.
  • I added JWT login, Apple Pay and Stripe payments, and push notifications, and worked with the founders and their clients the whole way through.
  • I moved the infrastructure from Akamai Linode to AWS with Terraform (EC2, S3, Route 53, IAM, Elastic Beanstalk, ElastiCache, load balancing) and wired in Amazon Bedrock.
Incuwise · 2024
03 Projects

Built on my own time, measured anyway.

01

Cost-Aware Semantic Gateway for LLM Routing

Sending every question to GPT-4o is expensive, and most questions don't need it. So I put a small classifier in front that guesses how hard each one is. Easy ones stay on Llama 3.2, running locally on vLLM, and only the hard ones go to GPT-4o.

68.1%lower cloud-model spend
79%of queries kept local
93.4%routing accuracy

The spend and share numbers come from an 81-query benchmark. Routing accuracy was measured on labeled prompts.

FastAPI · vLLM · Llama 3.2 · GPT-4o · Streamlit · Docker Source ↗
02

Automated LLM Evaluation & Observability Pipeline

I wanted a bad RAG change to fail the build, the same way a broken test does. Every pull request gets scored on faithfulness, answer relevancy, and contextual precision, and if any score drops below 95% the merge is blocked. With LangSmith tracing, working out why a check failed takes minutes instead of hours.

95%floor on every metric
3metrics per pull request
RAGAS · DeepEval · LangSmith · GitHub Actions · Claude Haiku Source ↗

04 Toolbox

Cloud & Containers

  • Kubernetes
  • Docker
  • OpenShift
  • AWS
  • Amazon Bedrock
  • Azure

Infrastructure & Delivery

  • Terraform
  • Kustomize
  • GitHub Actions
  • GitLab CI/CD
  • MLflow
  • Kubeflow
  • OAuth 2.0/OIDC
  • RBAC
  • HashiCorp Vault

AI Inference & Serving

  • vLLM
  • llm-d
  • Ollama
  • GPT-4o
  • Llama
  • Gemini
  • PyTorch
  • Hugging Face

Retrieval & Evaluation

  • RAG
  • GraphRAG
  • Hybrid search
  • BM25
  • Semantic chunking
  • Cohere reranking
  • LangChain
  • LangGraph
  • MCP
  • Tool calling
  • ReAct
  • RAGAS
  • DeepEval
  • LangSmith
  • LLM-as-a-Judge

Backend & Data

  • Python
  • SQL
  • JavaScript
  • Linux
  • FastAPI
  • PostgreSQL
  • Valkey
  • Redis
  • ElastiCache
  • Kafka
  • Weaviate
  • Neo4j

05 Education

Degrees

Stevens Institute of Technology

M.S., Machine Learning · Hoboken, NJ

Sep 2024 to May 2026

Central University of Haryana

B.Tech., Computer Science · India

Sep 2020 to Jun 2024

Certifications

  • AWS Certified Machine Learning Engineer, Associate2025
  • AWS Certified Developer, Associate2025
  • Databricks Certified Generative AI Engineer2026
  • Databricks Certified Associate Developer for Apache Spark2025
  • Anthropic Claude Certified Architect, Foundations2026
06 Contact

Want to work together? Say hello.

I'm looking for a full-time role in AI engineering, AI platform, MLOps, or forward deployed engineering. I'm based in New York and open to relocating. Email is the quickest way to reach me.

Rishi.Chhabra@outlook.com
Download resume ↓ GitHub ↗LinkedIn ↗Medium ↗