Kushal Kabra
Full-Stack AI Engineer
Kushal is a Full-Stack AI Engineer with practical expertise across applied machine learning, document intelligence, RAG systems, and production web development. He builds and ships end-to-end features combining modern AI/ML components with scalable, responsive full-stack architectures.
With hands-on experience in ML validation design, data leakage auditing, and large-scale data contract engineering across millions of production records, Kushal bridges robust AI modeling—including YOLOv8 layout classification, multi-engine OCR pipelines, and LLM fine-tuning—into dependable web applications.
What Kushal works on
Applied AI & RAG Architectures
Designs high-precision RAG systems, semantic retrieval engines, and LLM-powered conversational interfaces with structured context routing.
Document Intelligence & OCR
Builds multi-engine extraction pipelines combining custom YOLOv8 layout models, OpenCV preprocessing, and quality-score fallback mechanisms.
Full-Stack Web Engineering
Ships modern web platforms and advisory systems using Next.js 14, React, FastAPI, Node.js, and PostgreSQL with clean authentication and reactive state.
ML Validation & Data Engineering
Conducts rigorous data leakage audits, grouped train-test validation splits, and high-volume data warehouse queries across tens of millions of records.
Tools and technologies
Languages & Frameworks
- Python
- TypeScript
- JavaScript
- React.js
- Next.js
- Node.js
- FastAPI
- Express.js
- Tailwind CSS
- HTML5
- CSS3
Data & ML
- RAG Systems
- LLM Fine-Tuning
- YOLOv8
- Computer Vision
- NLP
- OpenCV
- LangChain
- FAISS
- HuggingFace Transformers
- PyTorch
- scikit-learn
- Gradient Boosting
- Random Forest
- pandas
Databases & Cloud
- Google BigQuery
- DuckDB
- PostgreSQL
- MongoDB
- MySQL
- Redis
- Docker
- AWS
- Vercel
- CI/CD
Validation & Tools
- ML Validation Design
- Data Leakage Auditing
- Git
- GitHub
- GitHub Pages
- REST APIs
- JWT Auth
Selected work
- 01
Full-Stack AI product engineering
Builds and ships product features across the entire software stack with applied AI/ML integrations, RAG pipelines, and responsive Next.js interfaces at Persist Ventures.
- 02
23% Precision@50 lift via ML benchmarking
Engineered baseline scoring rules and benchmarked Logistic Regression, Random Forest, and Gradient Boosting models, lifting Precision@50 from 0.62 to 0.76 on client-held-out splits at FlyRank.
- 03
Data leakage diagnosis on 79M-row warehouse
Identified critical validation leakage using GroupShuffleSplit to correct inflated metrics from 0.94 to an honest 0.76, auditing engineered features and querying 79M rows in BigQuery via DuckDB.
- 04
Universal multi-engine document extraction pipeline
Engineered an extraction pipeline across 10+ brands and 5+ invoice/document types, integrating Docling, PaddleOCR, TrOCR, EasyOCR, and Pytesseract with OpenCV preprocessing and automated fallback.
- 05
Custom YOLOv8 layout classification & RAG chatbot
Trained a custom YOLOv8 model to classify 7 document regions and constructed a natural-language Q&A chatbot using FAISS vector indexing and Groq LLM.
- 06
AI study-abroad advisory platform
Built a full-stack advisory application with Next.js 14, FastAPI, PostgreSQL, and Google Gemini API featuring real-time AI counselor chat, JWT authentication, and application tracking.
- 07
Ranked content action playbook & research publication
Authored and published a public research paper on GitHub Pages documenting machine learning methodology, reason-coded content action playbooks, and validation constraints.
