Jayesh Pandey

Jayesh Pandey

AI/ML Engineer at PGAGI | LLMs, RAG & Generative AI | Research Intern @ IEEE | Ex-ML Engineer @ PRL (ISRO) | AI Systems & Automation

Based in: Nashik, IndiaCurrently: AI/ML Intern, PGAGIExperience: 1 year

Hire me

A Computer Engineering senior with a proven track record of shipping production-grade ML systems and RAG architectures serving over 100K users. He specializes in translating complex AI capabilities into trustworthy workflows through rigorous offline evaluation and statistical modeling.

Experience

AI/ML Intern

PGAGIRemoteinternshipMay 2026 – Present

Summary

Developed and deployed Machine Learning and Generative AI API microservices on AWS EC2, handling 50K+ daily requests with sub-120ms latency and 99.9% reliability. Designed and implemented a Retrieval-Augmented Generation (RAG) pipeline using LangChain, FAISS, and Open Knowledge Format (OKF), improving retrieval precision by 40%, reducing hallucinations by 35%, and lowering query latency by 28%.

What I did

  • Translated broad business goals for an internal document RAG system into a testable technical pipeline including ingestion, chunking, embeddings, and evaluation.

Results

  • Handled 50K+ daily requests with sub-120ms latency and 99.9% reliability.
  • Improved retrieval precision by 40%, reduced hallucinations by 35%, and lowered query latency by 28%.
Machine LearningGenerative AIAPIAWS EC2RAGLangChainFAISSOpen Knowledge Format (OKF)Requirements GatheringStakeholder WorkshopsQLoRAQwenFine-tuning

Project Trainee - Data Analytics & ML

Physical Research Laboratory (PRL) / ISRORemoteinternshipFeb 2026 – Aug 2026

Summary

Engineered classical ML and statistical models (SciPy) to extract signal parameters from noisy satellite telemetry data, pushing extraction accuracy to 94.2% through rigorous offline evaluation. Architected automated data ingestion and preprocessing pipelines for IoT space weather sensors, accelerating dataset throughput by 3.2x and cutting analytical latency.

Results

  • Pushed extraction accuracy to 94.2%.
  • Accelerated dataset throughput by 3.2x.
Classical MLStatistical ModelsSciPyIoTData Ingestion Pipelines

Data & LLMOps Intern

EaseMeMedRemoteinternshipAug 2025 – 2026

Summary

Designed and iterated LLM prompts using few-shot and instruction-based techniques, then ran offline evaluation and benchmarking on model inference logs, cutting memory footprint by 45% and latency by 68%. Built dynamic data transformation pipelines and automated MLOps data augmentation workflows to accelerate classical classification model development, lifting multi-class classification mAP to 89.4%.

Results

  • Cut memory footprint by 45% and latency by 68%.
  • Lifted multi-class classification mAP to 89.4%.
LLMFew-shotInstruction-based promptingOffline evaluationBenchmarkingMLOpsData Augmentation

Skills

Technical

PythonPython
Retrieval-Augmented Generation (RAG)
Scikit-Learn
System Architecture Design
Offline Evaluation & Risk Assessment
Open Knowledge Format (OKF)
Prompt Engineering
LangChain
PyTorchPyTorch
Pandas
NumPy
Hypothesis Testing
FastAPIFastAPI
AWS (S3, EC2)
SQLSQL
FAISS
Requirements Gathering
PennyLane
Variational Quantum Circuits
DockerDocker
A/B Testing
Google Analytics
SEO
QLoRA
Model Evaluation
ReactReact
TypeScriptTypeScript
Fine-tuning
Google Search Console

Areas of expertise

Stakeholder Workshops

Projects

Askadoula.org

personal
A live full-stack platform where I worked on the frontend and backend integration.
ReactTypeScriptJavaScript

Codefy (WhoAmI)

personal
A developer security platform featuring a desktop application and an API backend.
TypeScript

RAG + Open Knowledge Format (OKF) Knowledge Engine

Lead Developerprofessional
Ran a system design phase before development - mapping data sources, defining a schema-consistent Open Knowledge Format (OKF), and stress-testing retrieval trade-offs - to ensure the architecture matched real query patterns rather than defaulting to free-text retrieval. Built an end-to-end RAG system normalizing heterogeneous data into OKF for schema-consistent vector indexing, improving answer grounding accuracy by 40% and cutting hallucination rate by 35% versus baseline free-text retrieval.
PythonLangChainFAISSOKFPyTorchGitHub

QHBERT - Quantum-Hybrid BERT for Fake News Detection

Researcher/Developerpersonal
Designed a hybrid quantum-classical transformer replacing a frozen DistilBERT encoder's classification head with an 8-qubit variational circuit, cutting trainable parameters ~2200x (110M to ~50K) and achieving a 98.85% F1-score across 4 fake-news datasets.
PyTorchPennyLaneDistilBERTMitiq
View the code

Ekip Bhaskar - Orbital Data Analytics Pipeline

Lead Developeracademic
Built a classical statistical anomaly detection model on Aditya-L1 solar payload telemetry, achieving 84.0% signal detection accuracy through iterative offline evaluation on raw space environment data streams.
PythonPandasSciPy

BTechNotes

personal
A platform for students to access notes and study resources, built during engineering years, which reached over 100,000 users.
Google Search ConsoleGoogle AnalyticsSEO

Domain-specific code/AI analysis fine-tuning

personal
Fine-tuned Qwen models using QLoRA for domain-specific code and AI analysis tasks.
QLoRAQwen

Education

Bachelor of Engineering (B.E.)

PVG's College of Engineering (PVGCOE SSDIOM)Computer EngineeringGPA 8.7Aug 2022 – Jun 2026

Recognition

Awards

NASA Space Apps Global Winner

NASA

Global winner for the Ekip Bhaskar project involving orbital data analytics.

2025 NASA Space Apps Global Award

Awarded for applying Earth-observation data to environmental challenges.

Certifications

Geodata Processing using Python and Machine Learning

ISRO / IIRS

Social impact & activities

Founder & Product Analytics Lead

BTechNotesorganizer

Grew and led a student-facing analytics platform to 100K+ active users and 1M+ views, translating statistical and correlation analysis of web telemetry into decisions non-technical stakeholders could act on.

Open Source Contributor & PyPI Maintainer

PyPIother

Authored PyPI Python data utilities FaceParser and TrainIQ; created 89+ public GitHub analytics & ML repositories.

Hire JayeshReach out about a role, a contract or a conversation.For recruiters

Jayesh's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis