Hi, I'm Navneet Singh

Full Stack Engineer focused on Applied AI Engineering. I build LLM applications, RAG systems, agent workflows, evaluation pipelines, and cloud-deployed AI products.

My work is mostly around taking AI ideas past the demo stage: retrieval, structured outputs, citation checks, evals, human review, rate limits, observability, Docker, and AWS/GCP deployment.

Current focus

• Applied AI engineering
• RAG, GraphRAG, and hybrid retrieval
• LangChain and LangGraph agents
• LLM evaluation and structured outputs
• Fine-tuning with LoRA and PEFT
• Human-in-the-loop AI workflows
• FastAPI, Next.js, Docker, AWS, and GCP

Tech stack

Languages

AI / LLM

Backend / frontend

Data / ML

Cloud / DevOps

Featured projects

CLARA - Credit Loan Analysis and Review Agent

Multi-agent loan review system built with FastAPI, Next.js, LangGraph, AWS Bedrock, Claude Haiku, Docker, and AWS Elastic Beanstalk.

• Built agents for term extraction, compliance checks, credit risk scoring, contradiction detection, counterfactual explanations, and human-review packet generation.
• Added audit logs, reviewer rationale, validated uploads, structured LLM errors, rate limits, and conservative fallback behavior.
• Improved final-outcome, compliance, and risk-band evaluation accuracy from 98% to 100% with zero regressions.

GitHub: https://github.com/navneet-singh2907/CLARA
Live: https://clara-web-beta.vercel.app/

FolioAware - Evidence-Grounded RAG Portfolio Assistant

Portfolio RAG assistant that answers only from owner-approved content, returns verified citations, and safely abstains when evidence is weak.

• Built with FastAPI, Vertex AI, Firestore vector search, Cloud Run, Terraform, and a TypeScript web component.
• Designed versioned knowledge sync with GitHub Actions and Workload Identity Federation.
• Created a 24-case evaluation suite with 100% Hit@5, 100% correct abstention, 100% citation precision, and 0% unsupported-answer rate on the accepted baseline.

GraphMind - AWS-Deployed Agentic RAG Engine

Agentic RAG engine for retrieving evidence across documents using vector search, BM25, metadata filtering, source inspection, and Neo4j graph traversal.

• Built a plan-verify-retry retrieval agent with evidence counts, sufficiency checks, confidence scores, attempts, and cited answers.
• Deployed to Amazon ECS Fargate behind an Application Load Balancer using Docker, ECR, Secrets Manager, CloudWatch Logs, and Neo4j Aura.

GitHub: https://github.com/navneet-singh2907/GraphMind

TicketRouter - Fine-Tuned Ticket Routing API

Fine-tuned Qwen3-1.7B with LoRA for 7-class IT support ticket routing.

• Improved validation accuracy from a 24.8% baseline to 77.8% and macro F1 to 0.753.
• Ran controlled A/B model comparisons across data changes, epoch count, class-level F1, and weak-class behavior.
• Deployed as a Dockerized FastAPI service on Google Cloud Run with Artifact Registry and a Streamlit operations console.

GitHub: https://github.com/navneet-singh2907/TicketRouter

AI-Powered Decision Intelligence Platform

Slack-based AI product that extracts business decisions from team conversations in near real time.

• Used LangChain structured outputs to identify decisions, owners, people, topics, and context.
• Modeled decision relationships in PostgreSQL with graph-style traversal.
• Built fail-closed owner detection, Slack correction flows, Docker deployment, and 250+ automated tests.
• Supported 10 active users across 4 monitored Slack channels.

GitHub: https://github.com/navneet-singh2907/convo-graph

pixelrag-langchain

Unofficial LangChain and LangGraph integration for PixelRAG.

• Built PixelRAGClient, PixelRAGRetriever, and PixelRAGSearchTool.
• Added screenshot-native visual retrieval, coordinate-based tile lookup, and optional inline base64 image support.
• Published to PyPI with 30 automated tests, Python 3.10-3.12 CI, and Trusted Publishing.

Earlier projects

RAG-Based AI Teaching Assistant

Built a local RAG assistant that converts video lectures into an interactive Q&A system using Whisper, embeddings, transcript chunking, and Streamlit.

Credit Card Fraud Detection Web App

Built and deployed a Django + scikit-learn fraud detection app with REST APIs, Isolation Forest anomaly detection, version-controlled ML artifacts, and Render deployment.

GitHub: https://github.com/navneet-singh2907/Fraud_Capstone

Professional experience

Full Stack Engineer (Remote) | Cryonex

Mar 2026 - Present

• Build and support customer-facing React and Next.js applications.
• Work across frontend, backend, production releases, bug fixes, and ongoing maintenance.
• Apply LLM-assisted development workflows to speed up implementation, debugging, testing, and delivery.

Software Developer | DITS

Jan 2019 - Dec 2020

• Built web applications using React, TypeScript, and Firebase.
• Implemented role-based authentication and Firestore-backed real-time data workflows.
• Worked with product teams to ship tested full-stack features.

Open source

I am building and contributing to open-source AI tooling, especially around RAG, agent workflows, and developer-facing LLM infrastructure.

pixelrag-langchain

Unofficial LangChain and LangGraph integration for PixelRAG.

• Built PixelRAGClient, PixelRAGRetriever, and PixelRAGSearchTool for screenshot-native visual retrieval.
• Added coordinate-based tile lookup, optional inline base64 image support, and self-hosted / hosted endpoint support.
• Published to PyPI with 30 automated tests, Python 3.10-3.12 CI, Trusted Publishing, and digital attestations.

OpenHuman

Contributed Korean internationalization support to OpenHuman, a privacy-focused agentic AI desktop application.

• Added locale translations and documentation fixes.
• Improved accessibility for Korean-speaking users.

Achievements

• Taught Python and web development to 50+ underprivileged students.
• Gold medalist in Boxing.

Connect

Email: [email protected]
LinkedIn: https://www.linkedin.com/in/navneet-singh2907/
Portfolio: https://navneetdatax.com
GitHub: https://github.com/navneet-singh2907