Faraz Ahmad

Associate AI Engineer | ML Researcher | Full-Stack AI Developer

Based in: Lahore, PakistanCurrently: Associate AI Engineer, Tkxel

Associate AI Engineer at Tkxel specializing in Computer Vision and LLMs. Proven track record in developing novel ML frameworks, including a geometry-aware knowledge distillation method that boosted MobileNet-V3 accuracy by 8%. Expert in fine-tuning VLMs and LLMs for specialized domains, with full-stack experience building AI-powered knowledge systems.

Experience

Associate AI Engineer

TkxelLahore, PakistanworkSep 2026 – Present

Built a custom person-retrieval dataset from 8 hours of client footage using YOLO cropping and ReID-based sampling; generated structured annotations with Qwen3.8-27B and manually verified a subset. Fine-tuned VLMs on the custom dataset & compared their performance gain over their vanilla counterparts. Created a diverse evaluation benchmark spanning images from different domains through crop curation, annotation verification, natural-language query generation, and query-to-image relevance mapping. Benchmarked re-ranking models and integrated the best-performing Qwen3-VL-Reranker-2B into the Smart Search pipeline to re-rank the top-k retrieved candidates.

YOLOReIDQwen3.8-27BVLMsQwen3-VL-Reranker-2BSmart Search

Engineering Intern, Computer Vision

TkxelLahore, PakistaninternshipJul 2026 – Aug 2026

Worked on a text-image person retrieval pipeline for natural-language person search over surveillance footage. Built a ≈ 99K-image pedestrian training dataset by mapping UPAR attribute annotations to PA-100K and generating diverse query-style text captions. Finetuned CLIP ViT-L/14, SigLIP2-so400m and IRRA ViT-B/16 using PyTorch and evaluated vanilla vs. fine-tuned models on PA-100K and real-world client data using mAP, MRR, Recall@K and Hit@K.

CLIP ViT-L/14SigLIP2-so400mIRRA ViT-B/16PyTorchPA-100KUPAR

Skills

Technical

PyTorch

Python

Finetuning

Hugging Face

CLIP

Supabase

YOLO

scikit-learn

NumPy

Next.js

HTML

Pandas

CSS

Google Colab

LLMs

RAG

Transformers

Kaggle

Git/GitHub

React.js

TypeScript

JavaScript

SQL

Matplotlib

Unsloth

Streamlit

CNNs

SigLIP2

Node.js

Express.js

REST APIs

PostgreSQL

Vector Databases

Vercel

Agentic workflows

Chroma

Docker

C/C++

Model Compression

TensorFlow

Azure

ONNX

Haskell

LoRA

Areas of expertise

Research

Projects

LLM Finetuning in Roman Urdu for Medical Domain (Applied ML)

personal

Built a custom medical dataset of ≈ 26k QA Roman Urdu pairs using gpt-4o-mini API for translation. Fine-tuned 1B, 3B, and 7B parameter LLMs and integrated a Hybrid RAG application with vector + keyword retrieval for domain-specific medical QA. Performed evaluation on the outputs using both automatic metrics (Bert-F1 score, chrF++) and human evaluation.

UnslothKaggleRAGStreamlitgpt-4o-mini APILoRA

View the code

Knowledge Distillation using Whitened-CLIP (Theoretical ML)

academic

Proposed and implemented a novel geometry-aware knowledge distillation framework to solve the 'double ellipsoid geometry' problem in CLIP models. By applying ZCA whitening to transform teacher embeddings into an isotropic hypersphere, the distillation process was significantly simplified for smaller, lower-capacity models. Verified the approach across multiple architectures, achieving an 8% accuracy boost on MobileNet-V3 and 5x faster convergence on ResNet-18.

PyTorchGoogle ColabKaggleHugging Face

View the code

AltBrain: Personal AI-powered Knowledge Base (Full-Stack AI)

Full-Stack AI Developerpersonal

Designed and deployed a markdown-first personal AI knowledge base for writing notes and asking AI questions. Implemented user-owned notes with Supabase Auth, Postgres, and Row Level Security, along with markdown preview, tags, wiki-style links, backlinks, daily notes, and search. Integrated Gemini-powered AI chat via a secure Next.js API route with two LLM modes: configured server key and bring-your-own Gemini API key.

Next.jsTypeScriptSupabasePostgreSQLGemini APIVercel

Visit the site View the code

Education

B.Sc. Computer Science

Lahore University of Management Sciences (LUMS)Computer ScienceGPA 3.56Aug 2022 – May 2026

Thesis. Geometry-Aware Knowledge Distillation using Whitened-CLIP

Coursework

Machine LearningDeep LearningAdvanced Topics in MLNatural Language ProcessingData ScienceDatabasesSoftware EngineeringCloud Development

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

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

More people on Kanvis

Browse more people Create your own Kanvis