Muhammad Sarmad Sohail

Who I Am

About Me

Passionate about leveraging data and AI to solve complex problems at scale

I'm a Data Scientist at Adept Tech Solutions, working on ASCEND, a synthetic data platform. I build its production data modules and the rule-based generation engine that layers business logic on top of them. Alongside that I research generative models for fraud detection, on financial, insurance and medical data where the fraud cases are a small fraction of the population. Most of the work turns out to be deciding whether the data you generated is any good, which is harder than generating it.

My background is Electrical Engineering at NUST, where I published at IEEE WCNC 2024 on deep reinforcement learning for IoT networks. I came into ML from analog circuit design: SAR ADCs in TSMC 28nm and GF 22nm FD-SOI, and a passive RFID transponder in CMOS 65nm. That is why I am drawn to ML applied to hard physical problems, from radar and signal processing to autonomy and instrumentation.

Outside work I build things end to end: a multi-agent job intelligence pipeline that runs unattended on a daily schedule, and Lantern, an open, model-agnostic coding agent. I was also named Top Performer at the Pak Angels Generative AI Program, Cohort 6, for work with BERT, LLaMA and LangChain.

Machine Learning

PyTorch TensorFlow scikit-learn GANs cVAE Autoregressive Deep RL

Data Engineering

Python PySpark Snowflake PostgreSQL SQL ETL Pipelines FastAPI

MLOps & Cloud

MLflow Optuna AWS Docker Kubernetes GitHub Actions

LLM & Agents

RAG LangChain PydanticAI FAISS Embeddings

Portfolio

Featured Projects

Synthetic data, generative models, agents, and IC design

Synthetic Data Generation

ASCEND Synthetic Data Platform

A platform that produces synthetic data for testing, development and analytics, from single flat files up to entire relational databases, without exposing real records. My work spans three connected tracks: the production data modules, the rule-based generation engine, and research on deep generative models.

Platform & data modules

Developed production-grade Python and FastAPI microservices for statistical analysis, preprocessing, data profiling, feature engineering, anonymization, curation, and synthetic data generation.
Built synthetic data generation module for relational databases sourced from PostgreSQL or Snowflake, producing privacy-safe versions of complete multi-table schemas or generating them later from extracted data signatures.
Implemented these capabilities as distributed PySpark jobs, orchestrating end-to-end ingestion, profiling, transformation, and generation workloads.

Rule-based generation engine

Built the platform's rule-based generation engine, layering configurable business logic over statistical generation through conditional rules within tables and cross-table rules across related data.
Implemented composite foreign-key resolution with per-parent cardinality defined as exact counts, ranges, or weighted distributions, keeping child records consistent with their corresponding count fields.
Developed the downstream Payload Expert tool to assemble nested API payloads from source-agnostic flat generated tables and produce deliberately malformed variants with required fields dropped, wrong types and invalid enums, for negative testing of a consumer API.

Enhanced synthetic data generation

Built and evaluated tabular synthetic data generators across three model families: GAN, autoregressive, and cVAE, for fraud detection on financial, insurance, and medical datasets under severe class imbalance.
Optimized model performance through Optuna-driven hyperparameter sweeps, MLflow experiment tracking, and GPU-based training on AWS instances.
Raised synthetic fidelity on a key transaction behaviour to 98.78% (98.96% for real data) by correcting the training feature set.

Python FastAPI PySpark PostgreSQL Snowflake SQL GANs cVAE Autoregressive Optuna MLflow AWS Docker Kubernetes

Proprietary

Deep Learning

CNN-LSTM for Video-Based Regression

Novel Conv2D + LSTM architectures for video-based regression on the UBFC dataset, with custom video preprocessing and data-augmentation pipelines.

PyTorch Keras CNN LSTM

Code

Multi-Agent

Job Intelligence Agent

A multi-agent system that tracks job opportunities, running unattended on a daily cron. Collectors pull listings from REST APIs, RSS feeds and HTML scraping, then a pipeline deduplicates them, embeds each posting with a sentence-transformer model, and ranks in two layers: cosine similarity against a target profile for recall, then LLM re-ranking over the shortlist. Top matches arrive as a daily notification.

Python LLMs Embeddings SQL GitHub Actions

NLP

Real-Time Meeting Summarizer

AI-powered transcription and summarization tool for meetings. Uses speech-to-text and LLM-based summarization for actionable insights.

Whisper LangChain Streamlit

Code

Top Performer

ZeroPhish Gate

Hybrid AI-powered phishing detection combining BERT and LLaMA for supply chain security. Built for the Pak Angels Generative AI Program.

BERT LLaMA LangChain Gradio

Code Certificate

Research Published

Resource Allocation in MEC-NOMA Networks

Deep reinforcement learning framework optimising computation offloading for energy-harvesting IoT devices. A DDPG agent dynamically balances time-sharing between data transmission and energy harvesting, minimising service delay and converging faster than baseline methods. Published at IEEE WCNC 2024.

Python TensorFlow Deep RL DDPG

Paper Demo Code Poster

Computer Vision

AI Construction Damage Analyzer

Vision-based chatbot for analyzing construction damage images using multimodal AI. Deployed as interactive Gradio application for real-time assessment.

Vision Models LLM Gradio

Code

GenAI

RAG System with Groq & FAISS

Retrieval-Augmented Generation system using Groq LLM and FAISS vector database for efficient document querying and context-aware responses.

LangChain Groq FAISS Gradio

Code

IC Design

Passive RFID Transponder Chip

Ultra-low-power RFID transponder for animal tagging in CMOS 65nm, built as the project for a five-month analog and mixed-signal training programme at the NUST Chip Design Centre. Under 10 µW power consumption and a 10 cm read range.

Cadence Virtuoso CMOS 65nm RFID

NCDC

AI Agents

Autonomous AI Agent Demo

Multi-agent system demonstrating autonomous decision-making and task execution. Implements ReAct framework for reasoning and action.

LangChain Agents Python

Code

Digital Design

4-bit Microprocessor & VGA Controller

Custom 4-bit microprocessor with a hierarchical datapath and control logic, plus VGA line-drawing algorithms implemented in Verilog for FPGA deployment.

Verilog FPGA Digital Logic

Academic

Building Intelligent
Data Systems

About Me

Machine Learning

Data Engineering

MLOps & Cloud

LLM & Agents

Featured Projects

ASCEND Synthetic Data Platform

Platform & data modules

Rule-based generation engine

Enhanced synthetic data generation

CNN-LSTM for Video-Based Regression

Job Intelligence Agent

Real-Time Meeting Summarizer

ZeroPhish Gate

Resource Allocation in MEC-NOMA Networks

AI Construction Damage Analyzer

RAG System with Groq & FAISS

Passive RFID Transponder Chip

Autonomous AI Agent Demo

4-bit Microprocessor & VGA Controller

Publications

Optimizing Resource Allocation in MEC-Enabled CR-NOMA-Assisted IoT Networks: A DRL-Driven Strategy

Let's Work Together

Building IntelligentData Systems