Muhammad Sarmad Saleem

Muhammad Sarmad Saleem

AI Engineer specializing in LLM Agents, RAG, and Computer Vision

Based in: Lahore, PakistanCurrently: AI Engineer, Adept Tech Solutions, NSTPExperience: 2 years

Hire me

Muhammad Sarmad Saleem is an experienced AI/ML Engineer with a strong background in building LLM-based agents, RAG systems, and deep learning models for satellite imagery. He has a proven track record of delivering high-impact research and scalable software solutions, including published work in glacier prediction and multispectral image classification.

Experience

AI Engineer

Adept Tech Solutions, NSTPIslamabadworkSep 2026 – Present

Summary

Working on the product team, building LLM-based agents that reason over database schemas and constraints to generate realistic synthetic data. Designing and evaluating agentic approaches that preserve the statistical properties of the source schema.

What I did

  • Compare synthetic and source data using KS statistics for numeric columns and Jensen–Shannon divergence for categorical distributions.
  • Evaluate relationship preservation by comparing correlation matrices and child-to-parent row ratios.
  • Monitor model collapse and data leakage by tracking duplicate rates and distance to the nearest real records.
  • Validate downstream utility of synthetic data using train-on-synthetic, test-on-real (TSTR) methodologies and classifier-based indistinguishability tests.
  • Optimized inference pipelines and resolved library and dependency conflicts.
  • Wrote custom preprocessing utilities when existing tools were insufficient.

Results

  • Implemented a multi-layered fidelity evaluation pipeline for synthetic data that enforces 100% schema compliance, including primary-key uniqueness and foreign-key integrity.
  • Reworked the data-processing and inference pipeline to reduce unnecessary computation and memory usage, significantly improving processing speed.
LLMAI AgentsDatabase SchemasKolmogorov–Smirnov (KS) statisticJensen–Shannon divergenceSynthetic Data GenerationData Validation

Research Assistant

Machine Vision & Intelligent Systems Lab, NUSTPakistanworkOct 2024 – Jul 2026
Developed deep learning models for multispectral satellite image classification, improving generalization across diverse geographic regions. Designed scalable PyTorch pipelines for data preprocessing, training, and model evaluation to support reproducible remote sensing research.
Deep LearningPyTorchRemote SensingComputer Vision

DAAD Research Intern (Fully Funded)

Robotics Research Lab, RPTU KaiserslauternGermanyinternshipJul 2025 – Aug 2025
Developed deep learning models for glacier extent and elevation prediction using satellite imagery and climate datasets. Built automated experimentation pipelines for efficient, reproducible training and presented findings to international researchers.
Deep LearningSatellite ImageryClimate Datasets

Skills

Technical

PythonPython
RAG
GitHubGitHub
REST APIs
Prompt Engineering
HuggingFace
FastAPIFastAPI
Transformers
GitGit
Computer Vision
Pandas
NumPy
AI Agents
PyTorchPyTorch
LangChain
WebSockets
JavaScriptJavaScript
TensorFlowTensorFlow
Scikit-learn
OpenCVOpenCV
LangGraph
LLM Fine-Tuning
FAISS
Pinecone
DockerDocker
MLflow
DVC
Node.jsNode.js
WebRTC
React.jsReact.js
Next.jsNext.js
PostgreSQLPostgreSQL
MySQLMySQL
MongoDBMongoDB
C++C++
n8n
Model Context Protocol (MCP)
AWS
FlaskFlask
DjangoDjango
Java
CC
Synthetic Data Evaluation
Hybrid Search
Metadata Filtering
Independent project creation
Reranking
Technical Troubleshooting

Projects

GlacioVision: Glacier Prediction Platform

Lead Developeracademic
Developed deep learning models for glacier extent segmentation and elevation prediction using satellite imagery and climate data. Built a full-stack platform for real-time visualization and forecasting, achieving an IoU of 0.87; results accepted for publication at IEEE FMLDS 2026.
Deep LearningSatellite ImageryClimate DataFull-stack

Attention-Based Road Scene Segmentation

Developeracademic
Developed a DeepLabV3+ model with attention mechanisms for semantic segmentation of urban road scenes. Built an end-to-end MLOps pipeline using MLflow and DVC for experiment tracking, versioning, and reproducible training, achieving 92% pixel-wise accuracy.
DeepLabV3+Attention MechanismsMLflowDVCMLOps

Crop Intelligence System

Developeracademic
Developed a CNN-LSTM model for crop classification using multispectral satellite imagery collected across multiple growing seasons. Processed spatial-temporal agricultural data and achieved 79% classification accuracy with improved cross-region generalization.
CNN-LSTMMultispectral Satellite ImagerySpatial-temporal data

LawBot

Developerpersonal
Built a RAG system using LangChain and LLMs to answer questions over legal documents. Implemented FAISS-based semantic search for source-grounded responses.
RAGLangChainLLMsFAISS

BankAssist: RAG and Fine-Tuned LLM Banking Chatbot

Developerpersonal
Built a RAG-powered banking assistant using FastAPI, Streamlit, and FAISS for intelligent customer support. Fine-tuned an LLM to provide accurate, domain-specific responses from enterprise documents.
RAGFastAPIStreamlitFAISSLLM Fine-Tuning

Education

BS Computer Science

National University of Sciences & Technology (NUST)Computer ScienceGPA 3.412022 – 2026

Recognition

Awards

DAAD Research Internship (Fully Funded)

DAAD2025

Fully funded research internship at RPTU Kaiserslautern, Germany

Patents & publications

GlacioVision: Deep Learning for Predicting Future Glacier Extent and Elevation Change from Remote Sensing and Glacio-Meteorological Data

publication2026

Accepted for IEEE International Conference on Future Machine Learning and Data Science (FMLDS)

With M. S. Saleem, et al.

Multi-Task Rank-Aware Learning for Stable Elite Genotype Selection from UAV-based Multispectral Temporal Imagery under Data Scarcity

publication2026

Accepted for IEEE International Conference on Future Machine Learning and Data Science (FMLDS)

With M. S. Saleem, Z. Mahmood, Z. Zafar, M. M. Fraz

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

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

More people on Kanvis

Browse more peopleCreate your own Kanvis