YP
Available
Currently · Associate AI Engineer @ Fx31 Labs

Building neural
systems that think,
reason, and ship.

I'm Yog Prajapati — an AI Engineer focused on building practical AI products using LLMs, RAG systems, and modern AI workflows. I work on transforming AI ideas into real-world applications by combining backend engineering, intelligent retrieval systems, and scalable product architecture. I have experience building production-ready AI platforms, full-stack applications, and AI-powered user experiences.

View selected work Download CV Get in touch
5+
AI & Full-Stack Projects Built
15+
Production-Ready Features & Integrations
AI Systems
Built with LLMs, RAG & Modern AI Workflows
~/yog/model.pt
forward · t=0.142s
loss · 0.0271
online · Ahmedabad, IN
UTC+5:30
role
AI
focus
LLMs · RAG · Eval
stack
Python · React · Next.js · LangChain · FastAPI
builds
AI Products · Full-Stack Platforms · Production Systems
PyTorch TensorFlow JAX Hugging Face LangChain Weights & Biases Ray vLLM Pinecone Weaviate Triton ONNX CUDA AWS GCP Modal PyTorch TensorFlow JAX Hugging Face LangChain Weights & Biases Ray vLLM Pinecone Weaviate Triton ONNX CUDA AWS GCP Modal
About

Engineer who treats models like software.

I started in research, drifted into systems, and now I sit happily at the seam. I care about evaluation as much as architecture and shipping as much as state-of-the-art.

Open to collabs
default activationGELU

“A model that ships on Tuesday beats a model that benchmarks on Friday.”

— a mentor, c. 2022
Modeling
/modeling
  • LLM fine-tuning (SFT, DPO)
  • Transformers
  • Diffusion
  • Multi-modal
Systems
/systems
  • Distributed training
  • vLLM / TGI serving
  • Vector retrieval
  • Quantization
MLOps
/mlops
  • Eval harnesses
  • Telemetry
  • CI for models
  • Cost / latency tuning
Backend
/backend
  • Python · Go
  • FastAPI
  • Postgres
  • Kafka · gRPC
Beyond the keyboard
Climber, slow-coffee enthusiast, occasional speaker.
Coffee/Day · 2.5Bouldering · V6
Selected work

Projects that earned their place in prod.

A small slice — open-source experiments, internal platforms, and one or two papers that became products.

All projects
PROJECT_01

TrueLead

AI-powered B2B sales intelligence and lead enrichment platform

AI
multi-agent lead intelligence
Google ADK · Multi-Agent AI · Asynq · Redis
live

Built a production-ready AI sales intelligence platform that transforms raw company lists into verified, enriched prospect profiles using a cost-optimized multi-agent pipeline. The system performs live company research, legitimacy verification, AI-generated prospect reports, and personalized outreach email generation with mandatory human approval before finalization. Features include cache-first instant company search, real-time enrichment tracking, resilient background job processing, secure JWT authentication, and production-grade reliability patterns such as retries, dead-letter queues, idempotent APIs, graceful degradation, and request cancellation.

GoGinGoogle ADKMongoDBRedisAsynqNext.js 16TypeScriptZustandTailwind CSS v4AxiosZodJWT AuthenticationServer-Sent Events (SSE)
Source Live demo
PROJECT_02

MultiModal RAG System

AI-powered multimodal retrieval and reasoning platform

RAG
multimodal retrieval
LangGraph · LangChain · Vector Retrieval
live

Built a production-ready multimodal RAG system using LangGraph and LangChain capable of processing programming languages, Github PRs, documentation, and structured data with intelligent retrieval, contextual reasoning, semantic search, conversation memory, and AI-powered response generation workflows.

PythonFastAPILangChainLangGraphFAISSMongoDBOpenAI APIs
Source Live demo
PROJECT_03

HireSight

AI-powered resume ATS analyzer and interview preparation platform

ATS
resume analysis
LLM · Resume Intelligence
live

Built an interview preparation platform featuring mock tests, AI-powered ATS resume analysis, skill assessments, and personalized feedback to help candidates improve job readiness and resume quality.

ReactNode.jsExpressMongoDBTailwind CSSAI/LLM APIs
Source Live demo
PROJECT_04

iPharm

AI-powered pharmacy and medicine recommendation platform

5000+
SKU support
ML · Recommendation Engine
live

Developed a scalable multi-vendor pharmacy platform with pincode-based medicine discovery, AI-powered recommendation engine, real-time user tracking, secure payments, and Socket.IO integration.

ReactNode.jsMongoDBPython MLRedisSocket.IO
Source Live demo
PROJECT_05

Leaf AI

Deep learning-based potato leaf disease detection system

CNN
disease classification
CNN · Image Classification
live

Created a real-time plant disease detection platform using Deep Learning and computer vision. Integrated image preprocessing, intelligent validation, and CNN-based classification for accurate disease prediction.

TensorFlowFastAPIPythonNext.jsOpenCV
Source Live demo
Neural Studio

The architectures I think in.

A visual index of the model families I work with daily — from the humble MLP to the transformer block that ate the world.

01 · SELF-ATTENTION
softmax(QKᵀ / √dₖ) · V

Scaled dot-product attention

Every token gets to query every other token. Heavy upfront — O(n²) — but parallelizable, and the inductive bias that unlocked LLMs.

queries
Q→ “the”
keys
K→ “model”
values
V→ “learns”
02 · MLP
y = σ(Wx + b)

Feed-forward, the workhorse

Stacked linear + nonlinearity. Old, simple, and 60% of the FLOPs in every transformer block.

03 · CNN
conv(x, W) + b, stride=1, pad=same

Convolutional feature maps

Local receptive fields, weight sharing, hierarchy. Still the right answer for pixels.

04 · RNN
hₜ = tanh(W·xₜ + U·hₜ₋₁ + b)

Recurrence over time

A single cell, reused. Beautiful in theory; you'll meet vanishing gradients on the way home.

05 · LSTM
C̃ₜ = tanh(W_c · [hₜ₋₁, xₜ])

Gates that remember

Forget, input, output. Add a carry line and suddenly long-range dependencies are tractable.

06 · TRANSFORMER
Vaswani et al., 2017

Block × N

Attention + FFN + residual + norm. Repeat until you have a personality.

07 · OPTIMIZATION
θ ← θ − η · ∇L(θ)

Gradient field

Backprop, one step at a time. Every arrow is a parameter that just got slightly less wrong.

08 · DECODING
temperature=0.7 · top_p=0.9

Token stream

What the model is doing right now, one piece at a time.

tokenize
embed
▁the
▁model
▁learns
→
logits[0]
0.124
softmax
▁to
▁predict
▁next
▁token
argmax
logits[1]
0.872
▁sample
temperature=0.7
▁top_p=0.9
▁repeat
EOS
kv_cache::hit
attn[0,12]=0.42
layer_norm
ffn
residual
→ output
tokenize
embed
▁the
▁model
▁learns
→
logits[0]
0.124
softmax
▁to
▁predict
▁next
▁token
argmax
logits[1]
0.872
▁sample
temperature=0.7
▁top_p=0.9
▁repeat
EOS
kv_cache::hit
attn[0,12]=0.42
layer_norm
ffn
residual
→ output
Experience · Education

A timeline, mostly linear.

Six years of work and learning. The interesting parts live in the bullets.

  1. Work· Jun 2026 — Present

    Associate AI Engineer

    Fx31 Labs
    • — developing and deploying RAG systems, retrieval pipelines, LLM powered applications, and scalable backend services for real world AI solutions
    • — Collaborating on production grade client projects and contributing to AI workflow automation and intelligent systems development.
    • — Handling client interactions in AI solution design and implementation.
    PythonFastAPIMicroservicesDockerRedisAgentic Workflows(Langchain/Langgraph/n8n)LLMsRAG
  2. Work· Dec 2025 — May 2026

    AI Intern

    Fx31 Labs
    • — Contributed to the development of RAG systems, retrieval pipelines, and LLM powered applications.
    • — Gained hands-on experience in building scalable backend services and AI workflow automation.
    RAGLLMsPythonFastAPIDocker
  3. Education· 2023 — 2026

    B.Tech Computer Science and Engineering (Specialization in AI/ML)

    GLS University
    • — Graduated with distinction. CGPA: 8.17/10.
    Data StructuresDeep LearningNLPComputer VisionReinforcement Learning
  4. Education· 2020 — 2023

    Diploma in Computer Engineering

    Gujarat Technological University
    • — Completed with distinction. CGPA: 7.92/10.
    Operating SystemsNetworkingDBMSTroubleshootingCompetitive Programming
Certifications · Achievements

Receipts.

Selected credentials and recognitions, in case those matter to you.

Cert· 2025
Create a Dashboard with Power BI
GLS University | Koenig Solutions | Microsoft
Cert· 2025
Green Skills & Artificial Intelligence Foundation
GLS University - Skills4Future Program
Cert· 2024
What is Generative AI?
LinkedIn
Cert· 2024
Programming with JavaScript
Meta via Coursera
Cert· 2023
Python for Data Science, AI & Development
IBM via Coursera
Cert· 2023
Fundamentals of Java Programming
Board Infinity via Coursera
Tech stack

Tools I reach for first.

Pragmatic over fashionable. I'll happily swap any of these out if the problem says so.

AI / ML
12
  • Python
  • TensorFlow
  • PyTorch
  • Scikit-learn
  • Pandas
  • NumPy
  • OpenCV
  • Deep Learning
  • Machine Learning
  • NLP
  • Recommendation Systems
  • Computer Vision
LLM & GenAI
11
  • LangChain
  • LangGraph
  • RAG
  • LLM Integration
  • Prompt Engineering
  • Gemini API
  • Azure OpenAI
  • AI Agents
  • Vector Search
  • Embeddings
  • AI Workflow Automation
Frontend
8
  • React.js
  • Next.js
  • TypeScript
  • JavaScript
  • Tailwind CSS
  • HTML
  • CSS
  • Socket.IO
Backend
9
  • Node.js
  • Express.js
  • FastAPI
  • REST APIs
  • Microservices
  • WebSockets
  • Authentication
  • API Integration
  • Real-time Systems
Database & Storage
7
  • MongoDB
  • PostgreSQL
  • MySQL
  • Redis
  • Supabase
  • Firebase
  • Cloudinary
Cloud / DevOps
9
  • Docker
  • Git
  • GitHub
  • Postman
  • Render
  • Vercel
  • Azure
  • CI/CD
  • Deployment
Tools & Platforms
8
  • VS Code
  • Cursor AI
  • Figma
  • Jupyter Notebook
  • Swagger
  • Linux
  • npm
  • GitHub Actions
Writing · Research

Notes from production.

Long-form posts, talks, and the occasional paper. Mostly things I wish I'd known earlier.

Full archive
EssayMay 2026· 12 min

Building a Production-Ready LangGraph Agent

State machines, tool nodes, and error recovery — everything the tutorials skip before you ship your first real agent.

Read
EssayMay 2026· 14 min

Multi-Agent Orchestration: Supervisor and Subagent Patterns

How to split complex tasks across specialised agents without losing control of state, context, or your sanity.

Read
EssayApr 2026· 10 min

Agentic RAG: Self-Correcting Retrieval That Works in Production

Naive vector search breaks silently. Corrective RAG loops catch the failure and rewrite the query before the user notices.

Read
EssayApr 2026· 11 min

Memory in LangGraph: Checkpointing, Persistence, and Long-Term Context

Three memory tiers — in-thread, cross-thread, and external store — and exactly when each one becomes the right tool.

Read
Contact

Let's build something useful.

Open to AI Engineering roles, applied AI collaborations, and building scalable AI-powered platforms using LLMs, RAG, and modern AI workflows. I usually respond within 48 hours.

Email
contact@yogprajapati.site
GitHub
@yogprajapati
LinkedIn
in/yogprajapati
Encrypted in transit · Never shared