AI/ML Engineer & Data Scientist

Production-focused AI/ML & GenAI Engineer specializing in engineering low-latency Generative AI architectures, real-time voice agents, and edge computer vision systems. Proven track record of architecting scalable enterprise solutions—from real-time speech-to-speech telephony and high-throughput LLM security gateways to commercially deployed biometric platforms.

Download CV LinkedIn

Core Technical Skills

AI & NLP

AI & NLP

Frameworks & Agents
LangChain LangGraph LlamaIndex Groq Pipecat Kokoro TTS VLM
LLMs & Multimodal AI
Gemini API Multi-modal AI RAG Architecture Imagen Ollama Qwen 2.5
Techniques & Guardrails
Prompt Engineering Fine-Tuning Chatbots Conversational Agents RAG Evaluation Guardrails AI
NLP & Deep Language Models
Transformers Attention Hugging Face NLTK spaCy TF-IDF Word Embeddings Text Summarization Sentiment Analysis Topic Modeling
Machine Learning

Machine Learning

Core Libraries
Pandas NumPy Scikit-Learn
Algorithms & Predictive Modeling
XGBoost LightGBM CatBoost Predictive Analytics Time Series Recommendation Systems EDA
Validation & Optimization
Cross-Validation Hyperparameter Tuning Grid Search Random Search Feature Selection Dimensionality Reduction
Data Visualization
Matplotlib Seaborn Plotly
Deep Learning

Deep Learning

Deep Learning Frameworks
PyTorch TensorFlow Keras Hugging Face
Neural Architectures
CNN RNN LSTM ANN GAN Autoencoders
Techniques & Generation
Transfer Learning Data Augmentation Image Processing Image Classification Text Generation
Computer Vision

Computer Vision

Vision Frameworks & Models
OpenCV YOLO MediaPipe OpenPose ArcFace (ResNet50)
Spatial & Motion AI
Object Detection Skeleton Tracking Angle Estimation
Biometrics & Facial AI
Face Recognition Face Verification Face Swap Emotion Detection Age & Gender Estimation VR180 Conversion
Python & Deployment

Python & Deployment

Web & Desktop Frameworks
Streamlit Gradio Flask Kivy FastAPI PyQt6
DevOps, APIs & Monitoring
API Deployment Monitoring Docker CI/CD
Web Scraping & Extraction
Data Scraping Beautiful Soup
Cloud & Version Control
AWS GCP SQL Git GitHub Heroku Netlify Hostinger
Frontend & Databases
HTML CSS JS MYSQL
Education & Languages

Ed & Languages

Academic Background
BCA (IGNOU) Computer Science & Applications
Languages & Proficiency
English (Professional Working) Hindi (Native / Fluent) Japanese (Intermediate N4/N5)
Career Journey & Impact

Professional Experience

Hands-on engineering across computer vision biometric pipelines, event-driven voice agents, predictive ML models, and production VPS deployments.

AI/ML & Data Science Specialist

Full-Time
Tips-G Alwar • Alwar, Rajasthan, India
01/2026 – Present

Leading the development and production deployment of AI-powered computer vision infrastructure, real-time conversational audio agents, and mentoring aspiring AI/ML engineers on industry-standard engineering practices.

TIPS-G AI Attendance System
Microsoft Store Official <100ms Inference Dockerized API
TIPS-G Attendance System on Microsoft Store
Publisher: Uvais Saifi Production Live

What I Built: I architected and built TIPS-G Campus Guardian, an enterprise edge-AI biometric attendance and anti-proxy system officially published on the Microsoft Store. Built for schools and colleges to eliminate attendance fraud and manual paperwork with real-time <100ms face recognition, zero-trust hardware machine binding, and automated parent communication.

Frontend GUI & Video Threading

Built with PyQt6 (v6.6+) using a custom dark industrial theme. Implemented a non-blocking OpenCV QThread video engine processing 1080p @ 30 FPS feeds, real-time bounding box overlays, and multi-tier role dashboards (Admin, Faculty, Student).

Backend Microservices & Async Server

Built with FastAPI (Python 3.10+) and Uvicorn ASGI. Designed REST verification endpoints (/api/attendance/verify) with JWT Bearer tokens and bidirectional WebSocket channels for real-time live dashboard broadcast.

AI/ML Models & Implementation Strategy
  • Face Detection & Alignment: Integrated RetinaFace (ResNet50 backbone) via ONNX Runtime for multi-angle face detection and 5-point affine transformation alignment to 112×112 px.
  • Biometric Feature Extraction: Used ArcFace (ResNet50) extracting 512D unit-normalized feature embeddings. Calibrated Cosine Distance threshold ≤0.50 delivering >99.2% accuracy.
  • Anti-Spoofing & Liveness: MediaPipe 468-point 3D Face Mesh calculating z-axis depth disparity (Δz > 0.015) between nose tip and cheekbones to defeat 2D photo prints and mobile screens, combined with Eye Aspect Ratio (EAR < 0.19) blink detection.
  • Adaptive Template Learning: Exponential Moving Average blending (α=0.12) automatically evolves stored embeddings when high-confidence matches (≥80%) occur, adapting to natural aging, glasses, and beards without re-enrollment.
  • Zero-Trust Anti-Proxy: Cryptographic SHA-256 Hardware Machine Fingerprint (Motherboard UUID + CPU ID + MAC) + Campus Wi-Fi ARP BSSID verification preventing remote check-ins.
Docker, Packaging & CI/CD: FastAPI backend containerized on python:3.10-slim on port 8003; automated pytest pipeline; client packaged via PyInstaller and Inno Setup (tipsg_installer.iss) published on the official Microsoft Store.
Docker + MSIX
PyQt6 GUI ONNX Runtime RetinaFace ArcFace (512D) MediaPipe 3D Mesh FastAPI Docker (python:3.10-slim) Supabase Twilio WhatsApp / SMTP
Get on Microsoft Store →
IndiaWalls AI Voice Calling Admin Control Center
Admin Console: ai.rajasthanwall.com 450ms Latency
IndiaWalls AI Voice Calling System
Exotel Telephony

What I Built: I engineered a full-duplex conversational Speech-to-Speech (S2S) voice telephony agent that answers inbound customer calls over PSTN phone lines for Indiawalls Infratech. It qualifies customer boundary wall requirements, calculates instant estimates, and schedules plant callbacks with an ear-to-ear latency of ~450ms.

Frontend & Backend Architecture:

Live administrative dashboard on ai.rajasthanwall.com built with modern web UI. Backend powered by FastAPI (Python 3.10) and Uvicorn ASGI handling bidirectional WebSockets (/exotel-ws) streaming 16-bit Linear PCM audio at 16kHz.

AI/ML Models & Implementation Strategy:
  • Offline STT: Sherpa-ONNX IndicConformer NeMo CTC model loaded as an in-memory singleton (~80ms latency, zero cloud API fees, zero quota limits).
  • Multi-Account Key Rotator: Google Gemini 2.0 Flash managed by a custom round-robin key rotator class with automatic 60s cooldown on 429 quota exhaustion + Groq LPU fallback (~150ms TTFT).
  • Neural TTS: In-process Kokoro-82M neural synthesizer (~150ms) streaming 16kHz audio chunks with acoustic echo suppression to prevent speakerphone feedback.
Docker & VPS Deployment: Multi-stage Docker container deployed on Cloud VPS (ai.rajasthanwall.com), Nginx reverse proxy with WSS termination, automated latency profiler, and Supabase/SQLite dual persistence.
Exotel PSTN IndicConformer STT Kokoro-82M TTS Gemini Rotator FastAPI Docker VPS
Admin Console →
RajasthanWall AI Assistant Live Chatbot
Live Hinglish Chatbot rajasthanwall.com
IndiaWalls Bilingual & Hinglish RAG
300ms Turnaround

What I Built: I built a production customer support AI assistant directly embedded on RajasthanWall.com. It accurately answers precast RCC boundary wall inquiries across English, Hindi, and colloquial Hinglish ("apke yaha konsi boundary wall banti hai") with 100% zero-hallucination guarantee.

Frontend & Framework-Free Strategy:

Custom vanilla JavaScript floating chat widget with suggestion chips. Built a framework-free pure Python backend bypassing LangChain/LlamaIndex bloat to achieve <650MB RAM on an economical Linux VPS.

AI/ML Models & Implementation Strategy:
  • Language Classifier: Heuristic regex detector classifying English, Devanagari Hindi, and Romanized Hinglish.
  • Embedding & Search: multilingual-e5-small (384D) + FAISS-CPU (IndexFlatIP) sub-millisecond semantic search.
  • Dual Guardrails: FAISS cosine cutoff (<0.70 dropped) + fast Groq Llama-3.1-8B fact check (~100ms) + Groq Llama-3.3-70B fluent synthesis with plant grounding (Kotkasim, Tapukara, Alwar).
Docker & VPS Host: Packaged via Docker Compose on Linux VPS hosting rajasthanwall.com, Nginx SSL reverse proxy, and automated CI/CD chunk tests (pytest).
FAISS-CPU multilingual-e5 Groq Llama-3.1 / 3.3 Docker Compose FastAPI
Live Website →
Technical Mentorship & Curriculum Direction

Mentored students and junior developers in PyTorch, OpenCV, Computer Vision pipelines, and production deployments.

AI/ML Engineer Intern

Internship
Syven • New Delhi, India
05/2025 – 12/2025

Architected production-ready predictive machine learning models for real estate valuation and designed semantic document intelligence RAG systems using LangChain and Google Gemini API.

Real Estate Price Prediction Model
~87% Accuracy

Developed a robust tabular price prediction engine trained on 10,000+ historical transaction records using XGBoost with domain-specific feature engineering.

  • Built automated preprocessing: missing value imputation, geospatial encoding, and outlier price normalization.
  • Evaluated performance with cross-validation and hyperparameter tuning to reach ~87% test accuracy.
XGBoost Scikit-Learn Feature Engineering Pandas
LLM Document Intelligence System
60% Faster Search

Designed and deployed an LLM-based intelligent search system utilizing LangChain and Gemini API for natural language extraction across 500+ legal and property documents.

  • Reduced document retrieval latency by ~60% compared to keyword-based search.
  • Implemented semantic chunking, prompt engineering, and context synthesis for high answer fidelity.
LangChain Gemini API RAG Vector Search Document Intelligence

Featured AI & Data Projects

Flagship Production Deployments (Tips-G & Client Systems)

Live in Production
TIPS-G Attendance System on Microsoft Store
Official Microsoft Store App <100ms Inference
Edge-AI Biometrics Zero-Trust HWID
TIPS-G Attendance System (Campus Guardian)

What I Built: I architected and built this enterprise-grade Edge-AI biometric facial recognition attendance platform published on the Microsoft Store. It replaces paper registers and eliminates proxy check-ins for colleges and universities with <100ms edge inference and zero cloud latency dependencies.

Frontend & Backend Architecture:

Developed in PyQt6 (v6.6+) with a multi-threaded OpenCV QThread video engine running 1080p @ 30 FPS. Backend built with FastAPI (Python 3.10) and Uvicorn ASGI with REST verification and WebSocket live dashboard dispatch.

AI/ML Models & Anti-Proxy Strategy:
  • RetinaFace + ArcFace (512D): Multi-angle face detection with 5-point landmark alignment via ONNX Runtime + ArcFace 512D unit embeddings (Cosine Distance ≤0.50, >99.2% accuracy).
  • 3D Anti-Spoofing: MediaPipe 468D Face Mesh evaluating z-depth disparity (Δz > 0.015) and EAR blink (<0.19) defeating 2D phone screens and photo prints.
  • Adaptive Template Learner (EMA): High-confidence matches automatically blend with stored template (α=0.12), adapting to aging and facial hair without re-registration.
  • Hardware HWID & Network Fencing: SHA-256 Motherboard/CPU hardware machine fingerprinting + Campus Wi-Fi ARP BSSID whitelist.
Docker, CI/CD & Cloud DB: Containerized on python:3.10-slim (port 8003); Supabase Cloud PostgreSQL with local encrypted SQLite fallback; automated pytest CI pipeline; packaged with PyInstaller & Inno Setup (tipsg_installer.iss) published on the official Microsoft Store.
PyQt6 GUI ONNX Runtime RetinaFace ArcFace (512D) FastAPI Docker Supabase
Get on Microsoft Store →
IndiaWalls AI Calling Admin Control Center
Live Admin Console 450ms Avg Latency
Speech-to-Speech (S2S) Exotel Telephony
IndiaWalls Real-Time AI Calling System

What I Built: I built an enterprise-grade Speech-to-Speech (S2S) conversational voice agent handling inbound and outbound phone calls over PSTN via Exotel for Indiawalls Infratech. It qualifies customer specifications, calculates pricing quotes, and schedules site inspections in real-time.

Architecture & Telephony Streaming:

Exotel IVR webhook returns Voice XML initiating bidirectional WebSockets (/exotel-ws) streaming 16-bit Linear PCM audio at 16kHz to an async FastAPI server. Built-in acoustic echo filter eliminates audio loopbacks.

AI/ML Models & Latency Strategy:
  • Offline STT: Sherpa-ONNX IndicConformer singleton running in-memory with zero cloud costs (~80ms decode).
  • Multi-Account Gemini Key Rotator: Google Gemini 2.0 Flash round-robin manager with automatic 60s cooldown on 429 quota errors + Groq LPU fallback (~150ms TTFT).
  • Neural TTS: In-process Kokoro-82M synthesizer streaming audio chunks to caller (~150ms). Total turnaround: 450ms.
Docker & VPS Telephony: Multi-stage Docker container deployed on Cloud VPS (ai.rajasthanwall.com) with Nginx WSS reverse proxy, live latency profiler, and Supabase PostgreSQL / SQLite dual persistence.
Exotel PSTN IndicConformer STT Kokoro-82M TTS Docker VPS Gemini Rotator WebSockets
Admin Control Center →
RajasthanWall AI Assistant Live Chatbot
Live Customer Assistant rajasthanwall.com
Live on RajasthanWall.com 300ms Turnaround
IndiaWalls Trilingual & Hinglish RAG Engine

What I Built: I built the official trilingual AI customer assistant on RajasthanWall.com. It accurately answers customer questions about precast RCC boundary wall pricing, height options, post dimensions, and factory logistics in English, Hindi, and romanized Hinglish ("bhai RCC wall kitne ki padegi").

Framework-Free Pure Python Architecture:

Completely bypassed heavyweight frameworks (LangChain / LlamaIndex) to eliminate memory overhead. Runs smoothly in <650MB RAM on an economical Linux VPS using FastAPI, sentence-transformers, FAISS-CPU, and SQLite.

AI/ML Models & Dual Anti-Hallucination Guardrails:
  • Language Detection & Embedding: Heuristic classifier routing language to multilingual-e5-small generating 384D dense vectors.
  • FAISS-CPU Vector Search: IndexFlatIP cosine similarity search over 14 precast boundary wall knowledge chunks.
  • Dual Guardrails: FAISS cosine threshold cutoff (<0.70 dropped) + fast Groq Llama-3.1-8B fact grounding check (~100ms) ensuring 100% zero-hallucination responses.
  • Synthesis: Groq Llama-3.3-70B fluent synthesis with regional factory grounding (Kotkasim, Tapukara, Alwar).
Docker & VPS Deployment: Docker Compose on Linux VPS hosting rajasthanwall.com, Nginx SSL reverse proxy, and automated pytest CI/CD pipeline validating chunk integrity.
Framework-Free Python multilingual-e5-small FAISS-CPU Docker Compose Groq Llama-3.1 / 3.3 FastAPI
Visit RajasthanWall.com →

Generative AI & RAG Prototypes

LLM Security Gateway Architecture & Dashboard
🛡️ AI Security Gateway Active Defense
LLM Security Gateway & Guardrails

An enterprise-grade, high-throughput security proxy protecting downstream Large Language Models (LLMs) against prompt injections, adversarial jailbreaks, and sensitive data leakage in real time.

▹ Threat Interception: Blocks semantic prompt injections, DAN/jailbreak vectors, and system prompt extractors before reaching inference.
▹ Automated PII Redaction: NER & regex pipeline masking credit cards, emails, SSNs, and phone numbers in both prompt & response streams.
▹ Token Rate Limiting & Proxying: Sliding-window token-bucket limiter with unified routing across OpenAI, Gemini, and open-source models.
FastAPI Python 3.10+ Guardrails AI Redis PII Masking Token Limiter
View GitHub → Security Proxy
Bill Image to ERP Automation System
⚡ VLM Document AI Edge Extension Store
Bill Image to ERP Automation System

End-to-end intelligent document processing pipeline that extracts invoice line items, tax breakdowns, and vendor metadata from bill images via Vision-Language Models (VLM) and directly populates ERP portals.

▹ Edge Browser Extension (Frontend): Published on Microsoft Edge Extension Store; features quick-capture HUD and automated DOM injection into ERP form fields.
▹ Cloud Backend on Render: Asynchronous high-performance FastAPI server deployed on Render Cloud handling secure multipart uploads and image preprocessing.
▹ VLM Multimodal Extraction: Vision-Language Model converts unstructured invoices into validated, schema-compliant JSON (GSTIN, line items, totals) with 98%+ precision.
Edge Extension Vision-Language Model (VLM) FastAPI Render Cloud ERP Sync Pydantic Schemas
View GitHub → Edge Store Ready
Mental Health AI App (Prototype)

Working RAG prototype utilizing the Gemini API and Imagen for enhanced mental health support. Focuses on contextual guidance and empathetic multi-modal response generation. (NLP and AI)

View GitHub →
RAG-based NLP Query Engine (Employee Data)

RAG system enabling employees to query structured data (CSV/Excel) using natural language. Leverages LLMs for semantic understanding and conversational inputs. (NLP and AI)

View GitHub →
Legal Document AI Extractor

An AI-powered application that extracts and structures key information from legal PDF documents using Google Gemini API. Features include executive summaries, purpose extraction, key definitions, eligibility criteria, obligations, and enforcement elements with a user-friendly Streamlit interface.

View GitHub →
AI Based Transcript Scoring System

An intelligent system that analyzes and scores students' spoken communication skills based on text transcripts and provided rubrics. Leverages Google Gemini API to evaluate multiple criteria including content knowledge, organization, language use, and delivery effectiveness.

View GitHub →
Iron Lady Chatbot

A Streamlit-based chatbot that answers questions about the Iron Lady organization using Google Generative AI (Gemini) embeddings and RAG (Retrieval-Augmented Generation). The chatbot retrieves relevant information from a predefined knowledge base and generates responses based only on that context.

View GitHub →
Dual Chatbot System Personal Assistant & Data Expert

A dual chatbot system built with Python, NLTK, TensorFlow/Keras, and a lightweight HTML+CSS+JavaScript front-end, designed to answer both personal and data‑science questions. The bots are trained on custom JSON intent files and are live on Heroku for instant access.

View GitHub →
Iron lady Simple App with AI Features

Iron Lady Simple App is an intuitive, AI-powered productivity tool built as a full-stack web application. It combines task management, course tracking, and feedback collection into a single, streamlined platform — all while leveraging generative AI to make suggestions, summarize user input, and help you stay organized.

View GitHub →

Computer Vision & Deep Learning

AthleteRise (Sports Performance Analysis)

Computer Vision tool using pose estimation (**Mediapipe**) to analyze and provide real-time feedback on athletic form and performance. (Computer Vision)

View GitHub →
Flexible Multi-Modal Emotion Detector

Modular classification system using Transformers and OpenCV to classify emotions from text, audio, or video inputs. Designed for high accuracy and multi-modal analysis in human-computer interaction. (Computer Vision)

View GitHub →
2D to 180VR Video Conversion Demo GIF
Convert 2D to 180VR Video

This project provides a pipeline to convert standard videos into VR180 SBS (Side-by-Side) format using MiDaS models (DPT_Large, DPT_Hybrid) for immersive VR experiences.It uses Gradio for a simple web interface and supports GPU acceleration (CUDA).

View GitHub →
COVID‑19 X‑Ray Classifier (Deployed)

This project provides an end-to-end solution for classifying COVID-19 from chest X-ray images using a Convolutional Neural Network (CNN) model. It includes a trained deep learning model and a user-friendly Streamlit web application for easy deployment and real-time predictions.

View GitHub →

Recommendation & Machine Learning & Predictive Modeling

Anime Recommendation System

This project implements an end-to-end Anime Recommendation System, leveraging machine learning to provide users with personalized anime suggestions. By analyzing anime genres, descriptions, and user ratings, the system employs Cosine Similarity and Natural Language Processing (NLP) techniques to deliver accurate and relevant recommendations.

View GitHub →
Stock Price Prediction Web App (Prophet)

Time series forecasting application using the Prophet library, deployed via Streamlit to offer interactive predictions. (Machine Learning)

View GitHub →
Dating Recommendation System

A hybrid recommendation model combining collaborative and content-based filtering to suggest optimal matches for users. (Machine Learning)

View GitHub →
Health Insurance Prediction (End-to-End)

A regression model predicting individual health insurance charges based on demographic and health factors, developed end-to-end. (Machine Learning)

View GitHub →
No Demo/Video Available
Detect Behavior with Sensor Data (Time Series)

A multi-class classification model analyzing time-series sensor data to accurately detect and categorize human behaviors. (Machine Learning)

View GitHub →
No Demo/Video Available
Hate Speech Detection (NLP Model)

This project provides an end-to-end solution for detecting hate speech in text data. It encompasses everything from data preprocessing and model training to deployment, making it a comprehensive toolkit for anyone looking to implement hate speech detection in their applications. The goal is to create a robust system that can identify and flag harmful content, contributing to safer online environments.

View GitHub →
NLP & Entity Extraction 89% Issue / 83% Urgency Acc
Customer Support Ticket Classification & Entity Extraction

Multi-label ticket classifier with TF-IDF and Logistic Regression on 1,000+ support tickets. Features custom sentiment polarity and text length engineering (+12% F1 boost) and regex-based entity extraction for 5+ entity types at 94% precision, deployed via Gradio.

TF-IDF Logistic Regression Gradio Regex Extraction
View on GitHub →

Data Scraping & Engineering Projects

Dataset & Scraper Kaggle Verified
Song Lyrics Scraper & NLP Dataset (2017–2024)

A web scraping project resulting in a comprehensive, quality-vetted NLP dataset of song lyrics from top artists, hosted on Kaggle for research and sentiment analysis.

BeautifulSoup Python Kaggle Dataset
View Dataset on Kaggle →

Problem Case

Solve the over and underfitting problem
RAG Evaluation Pipeline

A comprehensive pipeline for evaluating Retrieval-Augmented Generation (RAG) systems with automated metrics and analysis.

Python Code 
Loading...
Output Results 
Loading...
Input: Chat Conversation 
Loading...
Input: Context Vectors 
Loading...
View Full Project on GitHub →

Key Certifications

DeepLearning.AI & Coursera

Generative AI with Large Language Models

Focus: RAG, Fine-tuning, Deployment.

Private Course / Self-Paced

Python with Data Science

Core Python, Data Analysis, and ML fundamentals.

Udemy

Data Scientist BootCamp

Full-stack curriculum in Data Science.

SuperDataScience • Udemy

Machine Learning A-Z™ (Hands-On Python & R)

In-depth coverage of classical ML algorithms.

Coursera • Duke University

Managing Big Data with MySQL

Focus on efficient data querying and management.

Google Cloud Skills Boost

Explore Generative AI with the Gemini API

Hands-on application of Gemini for real-world tasks.

Google Cloud Skills Boost

Build Real-World AI Apps with Gemini & Imagen

Mastering multi-modal application development.

Google Cloud Skills Boost

Prompt Design in Vertex AI

Specializing in effective prompt engineering.

Google Cloud Skills Boost

Develop GenAI App with Gemini & Streamlit

Practical GenAI application development and deployment.

Google Cloud Skills Boost

Inspect Rich Documents with Gemini Multimodality & RAG

Advanced RAG for multi-modal document analysis.

Google Cloud Skills Boost

Natural Language Processing on Google Cloud

Practical NLP implementation on GCP.

Hack2Skill

Gen AI Academy

Comprehensive Generative AI training.

Ramjas College, Delhi University

Certification in Japanese Language N5

Foundation level Japanese proficiency.

St. Stephen's College, Delhi University

Diploma in Japanese Language N4

Intermediate level Japanese proficiency.

Let's Connect and Collaborate

I am actively seeking roles where I can apply my expertise in Generative AI and end-to-end Machine Learning solutions.

Email: uvaissaifi523@gmail.com GitHub Profile