Skip to content
View asghar-rizvi's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report asghar-rizvi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
asghar-rizvi/README.md
╔═══════════════════════════════════════════════════════════╗
║         ASGHAR QAMBER RIZVI  ·  AI/ML ENGINEER            ║
║         Building systems that think, remember, speak      ║
╚═══════════════════════════════════════════════════════════╝

LinkedIn Email GitHub


Who I Am

Python AI engineer who builds end-to-end intelligent systems — from fine-tuning LLMs and designing RAG pipelines to deploying production FastAPI backends that handle real users. I care about things that actually work: low latency, correct retrieval, reliable APIs, and models that behave in production, not just in notebooks.

Currently finishing CS @ Bahria University (3.81 GPA, graduating June 2026). Based in Karachi, Pakistan. Available for remote roles.


What I've Built

🔷 JUVO — Agentic AI Service Orchestrator

Python · FastAPI · Gemini AI · PostgreSQL/PostGIS · JWT · Render · Supabase

An agentic platform for Pakistan's informal economy. Users describe a service need in natural language (Urdu, Roman Urdu, or English) — JUVO's multi-agent pipeline extracts intent, finds geographically nearby providers via PostGIS spatial queries, and creates a confirmed booking, all within a single conversational flow.

  • Designed a multi-agent architecture: Intent Agent → Discovery Agent → Booking Service
  • Built Hold-to-Lock (HTL) reservation system with 5-minute expiry and ACID database triggers to eliminate 100% of double-booking conflicts
  • Implemented role-based JWT auth (user + provider), rate limiting, background task cleanup, and full Swagger documentation
  • Deployed on Render + Supabase with Docker; full provider dashboard with analytics

🔷 Dextera — AI-Powered Criminal Law Assistant

Python · LangChain · Llama-3 (3B) · LoRA · MongoDB Vector Atlas · ChromaDB · RAG

Domain-specific legal AI assistant built on fine-tuned Llama-3 with a production-grade RAG pipeline.

  • Fine-tuned Llama-3 (3B) using LoRA (r=32, α=64) on criminal law cases and legal statutes for improved legal reasoning
  • Engineered a vector database on MongoDB Vector Atlas — 0.98 accuracy on past cases and legal statutes
  • Production-level techniques: streaming responses, caching, rate limiting — query speed reduced to 200ms
  • Chat and message management via LangChain memory components

🔷 Real-Time Exercise Detection & Feedback System

Python · FastAPI · WebSocket · Computer Vision · Pose Estimation

Real-time fitness AI built during internship at Meta Frolic Labs.

  • Custom computer vision pipeline with pose estimation algorithms — 98% accuracy in live exercise detection
  • FastAPI backend with WebSocket integration for real-time data streaming at 0.3ms latency
  • Performance analysis algorithm evaluating form quality and delivering corrective suggestions through a web interface — 40% improvement in feedback system effectiveness

🔷 Urdu TTS with Voice Cloning — Zia Mohyeddin

Python · SpeechT5 · Transformer · Multi-speaker TTS

Speech synthesis system targeting the voice of Zia Mohyeddin, Pakistan's most celebrated literary narrator.

  • Developed SpeechT5 transformer-based architecture with multi-speaker capabilities
  • 25% improvement in naturalness over existing Urdu TTS solutions
  • 92% user preference in blind listening tests for voice fidelity

🔷 Emotion Conversion — GAN + Transformer Hybrid (R&D)

Python · PyTorch · GANs · Transformers · Audio Processing

Research project: converting emotional tone in speech while preserving semantic content.

  • Designed a hybrid GAN + Transformer architecture — Transformer vector embeddings carry semantic context; GAN learns the target emotion tone
  • Achieved 62% success rate on neutral → happy emotion conversion
  • Custom approach combining contextual sentence embeddings with adversarial training for audio emotion transfer

🔷 Paraphrase Model — Fine-Tuned LLM

Python · Transformers · Fine-tuning · NLP

Fine-tuned a language model specifically for high-quality paraphrase generation — preserving meaning while altering structure and vocabulary. Trained and evaluated on custom paraphrase datasets.


Stack

LLM & AI       Llama-3 · SpeechT5 · LoRA fine-tuning · GANs · Transformers
                Prompt engineering · RAG pipelines · LangChain · ChromaDB
                MongoDB Vector Atlas · Vector search · Embedding models

Computer       Pose estimation · Real-time inference · WebSocket streaming
Vision         OpenCV · MediaPipe

Backend        FastAPI · Django · Flask · SQLAlchemy · Pydantic
               JWT auth · Rate limiting · Background tasks · REST APIs

Databases      PostgreSQL · PostGIS · MongoDB · Vector databases · Redis

MLOps &        Docker · Render · Supabase · Alembic · Gunicorn · Nginx
Deploy         Multi-user concurrency · Production API design

Languages      Python (primary) · SQL · C++ · Java

By The Numbers

Metric Value
Exercise detection accuracy 98%
API latency (WebSocket, real-time CV) 0.3ms
Legal RAG retrieval accuracy 0.98
Query speed (RAG + streaming) 200ms
Urdu TTS improvement over baseline +25%
Voice clone user preference (blind test) 92%
Emotion conversion success rate 62%
Feedback system effectiveness improvement +40%
GPA 3.81 / 4.0

Experience

Python AI Trainee — Meta Frolic Labs (Aug 2025 – Oct 2025) Shipped real-world AI systems under senior engineers: real-time computer vision pipeline, emotion conversion research (GAN + Transformer), and Urdu TTS with voice cloning. Production-level work with FastAPI, WebSocket, and transformer fine-tuning.


GitHub Stats


Open to remote AI/ML engineering roles · UTC+5 · async-first · asgharqamberrozvi@gmail.com

Pinned Loading

  1. -Urdu-Text-to-Speech-with-Voice-Cloning -Urdu-Text-to-Speech-with-Voice-Cloning Public

    A fine-tuned SpeechT5 Urdu TTS model with voice cloning that converts both Urdu and Roman Urdu text into natural speech. Trained on diverse Urdu and Zia Mohiuddin recordings, it offers expressive, …

    CSS 5

  2. Exercise-Detector Exercise-Detector Public

    Exercise Detector is a real-time system that recognizes 7 different exercises using pose estimation and deep learning. It analyzes exercise form, counts repetitions, and achieves 99% accuracy on re…

    Jupyter Notebook

  3. Dextera-training Dextera-training Public

    Python

  4. AI-Lawyer-with-GUI AI-Lawyer-with-GUI Public

    Developed a legal AI chatbot using a custom LLM to assist lawyers by reading and summarizing legal PDFs. Built a Python GUI for seamless interaction and integrated LLMMA for efficient processing on…

    Python

  5. English-Paraphraser-FineTunned-LLM English-Paraphraser-FineTunned-LLM Public

    **PARA-PHRASER** is a fine-tuned **T5-small model with LoRA** designed for efficient and high-quality paraphrasing. It delivers **fluent, meaning-preserving outputs** while running smoothly on **CP…

    Jupyter Notebook

  6. Deep_Learning_Projects Deep_Learning_Projects Public

    Developed multiple Deep Learning projects using CNN, RNN, ANN, and fine-tuned models to tackle real-world challenges. Optimized architectures for diverse tasks, ensuring high accuracy and efficienc…

    Jupyter Notebook 1 1