class ManishJoshi:
role = "Software Engineer → AI/ML Engineer"
education = "B.Tech Computer Science & Engineering | 2025"
currently_learning = [
"Machine Learning",
"Deep Learning",
"Generative AI",
"RAG",
"MLOps"
]
building = [
"End-to-End ML Systems",
"Deep Learning Applications",
"RAG & LLM Applications",
"Production AI APIs"
]
philosophy = "Don't just train models. Build systems."My journey started with software engineering.
I worked on production applications where I learned how to build features, integrate APIs, work with backend systems, debug real problems and ship software that people actually use.
At Esferasoft Solutions, I worked on an AI-powered fitness application as well as a real-time laundry pickup and delivery platform.
Working around AI-powered features made me want to understand the technology behind them.
So I started going deeper into Machine Learning, Deep Learning, Generative AI and MLOps.
Instead of only following courses, I started building complete systems myself.
Today, my focus is on combining my software engineering foundation with hands-on AI/ML development.
Software Engineering
│
▼
Production Applications
│
▼
AI-Integrated Applications
│
▼
Machine Learning
│
▼
Deep Learning
│
▼
Generative AI
│
▼
MLOps & Deployment
│
▼
AI/ML Engineering
| 🧠 Machine Learning | 🔥 Deep Learning | 🤖 Generative AI | ⚙️ MLOps |
|---|---|---|---|
| XGBoost | PyTorch | RAG | Docker |
| Scikit-learn | CNNs | FAISS | MLflow |
| Feature Engineering | ResNet | BM25 | GitHub Actions |
| Optuna | YOLOv8 | Cross-Encoder | AWS EC2 |
| Model Evaluation | Transformers | NL-to-SQL | Monitoring |
Turning structured business data into answers through natural language.
Dataset
│
▼
DuckDB Ingestion
│
▼
Auto Profiling
│
▼
┌────────────────────────┐
│ Hybrid Retrieval │
│ │
│ FAISS Dense Search │
│ + │
│ BM25 Sparse Search │
│ ↓ │
│ RRF Fusion │
│ ↓ │
│ Cross-Encoder Ranking │
└────────────────────────┘
│
▼
NL-to-SQL + RAG
│
▼
LLM
│
▼
Business Insight
- ⚡ Processes 60K+ row datasets in under 0.3 seconds
- 🚀 Reduced BI turnaround from hours to under 10 seconds
- 🎯 Achieved 98% hit-rate@5 on retrieval quality
- 🔎 Hybrid FAISS + BM25 retrieval
- 🔀 RRF fusion
- 🧠 Cross-encoder reranking
- 💬 Natural-language data querying
- 🌊 SSE streaming
- 🤖 Groq Llama 3.3 70B
- 💾 Designed to operate within a 256 MB memory budget
A machine learning project focused on both model performance and reliable deployment.
Raw Dataset
│
▼
Data Validation
│
▼
Feature Engineering
│
▼
Class Imbalance Handling
│
▼
XGBoost
│
▼
Optuna Hyperparameter Tuning
│
▼
Model Evaluation
Git Push
│
▼
GitHub Actions
│
▼
Docker Build
│
▼
Validation + Tests
│
▼
AWS EC2
│
▼
MLflow Tracking / Registry
- Dataset imbalance: 73:27
- XGBoost classifier
- Optuna hyperparameter tuning
- Class-weight balancing
- Recall increased to 82%
- Great Expectations data validation
- MLflow experiment tracking
- MLflow model registry
- Training-serving skew safeguards
Before → 25 minutes
After → < 3 minutes
~90% reduction in deployment time per push.
From raw environmental audio to an explainable deep-learning prediction.
🎵 Raw Audio
│
▼
📊 Mel Spectrogram
│
▼
🧠 ResNet-34 Style CNN
│
▼
🎯 50 Environmental Sound Classes
│
▼
🚀 FastAPI + Modal
│
▼
🌐 React Visualization
- Custom ResNet-34-style 2D CNN
- 21M+ parameters
- 50 environmental sound classes
- PyTorch
- torchaudio
- Mel-spectrogram input
To address overfitting on a small training dataset:
- SpecAugment
- Mixup
- Label smoothing
Optimizer → AdamW
Scheduler → OneCycleLR
Training → 100 epochs
GPU → NVIDIA T4
Framework → PyTorch
The model is deployed through a serverless FastAPI endpoint on Modal with a React frontend for:
- Audio waveform visualization
- Predictions
- CNN activation maps
- Explainable model output
Mohali, India · Sept 2025 – April 2026
My professional experience has been focused on building production software while working closely with AI-powered features.
Worked on:
- AI-generated workout plans
- AI-generated meal plans
- Body scan flows
- Food scan flows
- Conversational fitness chatbot
- Backend-hosted LLM integration
- REST API integration
Worked on:
- Driver application
- Customer application
- Google Maps integration
- WebSockets
- Real-time driver location tracking
- Backend integration
- Production feature development
- Debugging
- Code reviews
- Clean and maintainable code
- Collaboration with backend teams
OOP Data Structures & Algorithms
Feature Engineering
Hyperparameter Tuning
Optuna
Model Evaluation
Statistical Analysis
EDA
Matplotlib Seaborn Plotly
CNNs ResNet YOLOv8 Transformers
RAG
Hybrid Dense + Sparse Retrieval
Cross-Encoder Reranking
NL-to-SQL
Vector Databases
OpenAI API Gemini API Groq API
FastAPI Docker CI/CD AWS EC2
MLflow Great Expectations Modal
Model Monitoring
PostgreSQL DuckDB MongoDB Redis
REST APIs WebSockets Git
Amrapali Institute of Technology, Uttarakhand
August 2021 → June 2025
Artificial Intelligence
Machine Learning
Deep Learning
Probability & Statistics
Data Structures & Algorithms
Object-Oriented Programming
System Design
Top 1.5% Nationwide
98.48 Percentile
162+ Problems Solved
Dynamic Programming · Graphs · Monotonic Stack · Sliding Window
I don't see an ML project as just:
Dataset → Model → Accuracy
I think about the complete system:
┌──────────────┐
│ DATA │
└──────┬───────┘
↓
┌──────────────┐
│ PREPROCESSING│
└──────┬───────┘
↓
┌──────────────┐
│ MODEL │
└──────┬───────┘
↓
┌──────────────┐
│ EVALUATION │
└──────┬───────┘
↓
┌──────────────┐
│ API │
└──────┬───────┘
↓
┌──────────────┐
│ DEPLOYMENT │
└──────┬───────┘
↓
┌──────────────┐
│ MONITORING │
└──────┬───────┘
↓
┌──────────────┐
│ PRODUCT │
└──────────────┘
That's the direction I'm working toward:
AI / ML
│
┌──────────────┼──────────────┐
↓ ↓ ↓
Machine Learning Deep Learning Generative AI
│ │ │
XGBoost PyTorch RAG
│ │ │
Evaluation CNNs LLM APIs
│ │ │
└──────────────┼──────────────┘
↓
MLOps
│
Docker / MLflow
│
FastAPI
│
AWS EC2
↓
AI/ML ENGINEERING
If you're working on AI/ML, Generative AI, Deep Learning, RAG, MLOps or AI products, I'd love to connect.
Learning → Building → Deploying → Improving