~/rishon-jain

00 · init — DATA SCIENCE · ML · AI ENGINEERING

I turn messy data into shipped models.

I build end-to-end ML systems, from data and training through evaluation and deployment. Data Science Intern at a GovTech AI startup.

0.9907
AUC-ROC on MULTITuDE
88K
Wikipedia articles in one RAG system
78%
mAP@0.5 on road damage
k = 3 · iter 12 · inertia 0.184● cursor = data point

01 · ingest — about

Building ML from data to deployment.

I build end-to-end ML systems, from data and training through evaluation and deployment, with a focus on computer vision and LLM applications.

I'm a Data Science Intern at Blitz AI, working on AI-driven plan review for local governments. Alongside that, I build local-first ML systems: a hallucination-checked RAG chatbot, a multilingual AI-text detector with an accompanying paper, and a road damage detector.

Portrait of Rishon Jain

$ now --status

buildingcomputer vision for plan review
readingLLM inference cost at scale
learningLangGraph and evals
locationNoida, UP

02 · transform — experience

Experience, as a git log.

  1. b7e41a2
    JUL 2026 — PRESENT

    Data Science Intern · Blitz AI

    Data science intern on the AI Pre-check and Plan Review Platform that helps local governments review building plans.

    • Build and fine-tune computer vision models that read architectural and construction plan sets.
    • Work on how LLM and vision-language components fit into the review pipeline, with a focus on inference cost at scale.

    Described at a high level to respect confidentiality.

  2. 4d09c6f
    JUL — SEP 2025

    AI Content Research Intern · Guvi HCL

    Researched 30+ AI and ML topics and turned them into accessible content. Used Google Trends and YouTube Analytics to find high-impact topics and worked across engineering and content teams.

03 · model — projects

Projects as model cards.

Model card · 01GitHub ↗

Hallucination-Aware RAG System

A fully local, self-correcting RAG chatbot over 88K Wikipedia articles. A DeBERTa NLI checker verifies every answer against retrieved context and triggers up to 3 LangGraph retries before replying.


data
88K articles · 150K+ chunks
metric
Grounded answers, offline
stack
LangChain · LangGraph · FAISS · Ollama · Streamlit
limits
Runs Llama 3.2 3B on 4GB VRAM, so answers are only as strong as that model.
Model card · 02GitHub ↗

PolyDetect

A hybrid detector for machine-generated text across languages. Frozen transformer embeddings are combined with perplexity and lexical-diversity features.


data
MULTITuDE benchmark
metric
AUC-ROC 0.9907 (MiniLM + XGBoost)
stack
PyTorch · TensorFlow · Hugging Face · XGBoost
limits
Tuned for a 4GB single workstation; results come from one benchmark.
Read the paper ↗
Model card · 03GitHub ↗

RoadSense

End-to-end road damage detection with YOLOv8n and a Streamlit dashboard. It takes video uploads, runs async inference and generates annotated PDF reports with severity classes.


data
RDD2022 · 4 damage classes
metric
78% mAP@0.5 · 30+ FPS on GPU
stack
YOLOv8 · OpenCV · Ultralytics · Streamlit
limits
Four damage classes only, trained on RDD2022.

04 · evaluate — skills

DATA ENGINEERING

  • PostgreSQL
  • PySpark
  • Kafka
  • Docker
  • Kubernetes
  • AWS
  • Azure
  • Git
  • CI/CD

ANALYTICS

  • Python
  • SQL
  • Pandas
  • NumPy
  • Power BI
  • Tableau
  • Streamlit

MACHINE LEARNING

  • PyTorch
  • YOLO
  • scikit-learn
  • Hugging Face
  • TensorFlow
  • OpenCV
  • XGBoost

AI ENGINEERING

  • LangGraph
  • LangChain
  • FAISS
  • Ollama
  • RAG
  • NLI hallucination checks
  • REST APIs

05 · validate — certifications

06 · serve — writing (substack)

07 · deploy — contact

Let’s build something that ships.

>ESC
↑↓ navigate↵ selectesc close