Mohib Ullah Khan Sherwani

Full-stack AI Engineer & Gold Medalist specializing in production-grade ML and RAG systems.

Based in: Lahore, PakistanCurrently: Full-stack AI Engineer, Cognilium AIExperience: 1 year

Hire me

A high-achieving software engineer who built one of the largest public Pakistan Sign Language datasets and trained recognition models to 92% accuracy. He has extensive experience shipping end-to-end AI solutions using AWS, Azure, and multi-agent LLM frameworks to drive significant operational efficiency.

Experience

Full-stack AI Engineer

Cognilium AILahore (On-site)workAug 2026 – Present

Summary

Replaced manual website data collection with a Slack command, as measured by 3 sites scraped concurrently with progress tracking and retries, by building it solo on Google ADK with SQS, ECS Fargate, DynamoDB, S3, SNS, and Lambda defined in AWS CDK. Delivered classroom-ready lessons in under 12 seconds, as measured by rubric-validated output over 584 curated passages and a 298-term vocabulary, by implementing client-requested features in a FastAPI, Qdrant, hybrid RAG pipeline.

Results

  • Enabled concurrent scraping of 3 sites with progress tracking and retries.
  • Reduced lesson delivery time to under 12 seconds for 584 curated passages.
Google ADKSQSECS FargateDynamoDBS3SNSLambdaAWS CDKFastAPIQdrantRAG

AI Engineer

KalSoftRemoteworkMay 2025 – Aug 2026

Summary

Cut forecasting time 70% across 750K+ sales records and 50+ products, by building warehouse-level LightGBM models with drift prevention and CI/CD on Azure DevOps and GitHub. Cut manual review effort 80% across 50+ applications per job, by building a 9-agent recruitment platform with deterministic filtering ahead of LLM matching and strict output validation. Enabled live voice-based candidate interviews by integrating Whisper speech-to-text and Azure TTS into the recruitment pipeline. Cut contract evaluation from 4 hours to 10 minutes (96%), by building an async RAG risk-scoring agent over 4 data silos with response validation and retries. Cut reporting time 70% for non-technical teams, by building a validated multi-agent text-to-SQL system with safe execution and automated charts.

What I did

  • Developed a pre-LLM deterministic pipeline that calculates a candidate's matching percentage against a job description.
  • Developed a demand forecasting platform where a model trained on 750k records was integrated as a tool for an agent via a simplified .predict function interface.

Results

  • Optimized token costs by building a pre-LLM deterministic pipeline to match candidate skills and experience before LLM processing.
  • Ensured strict output validation and prevented prompt bloating or hallucinations by implementing Pydantic.
  • Cut forecasting time by 70% for 750K+ sales records.
  • Reduced manual review effort by 80% using a 9-agent recruitment platform.
  • Reduced contract evaluation time from 4 hours to 10 minutes (96% reduction).
  • Cut reporting time by 70% for non-technical teams via text-to-SQL system.
  • Handled 50+ applicants per role for PHDC Ghana using a cost-optimized processing pipeline.
LightGBMAzure DevOpsGitHubWhisperAzure TTSRAGText-to-SQLPydantic

Skills

Technical

PythonPython
TypeScriptTypeScript
RAG
TensorFlowTensorFlow
FastAPIFastAPI
MediaPipe
AWS
ReactReact
LangChain
Google ADK
Qdrant
DockerDocker
Azure
PostgreSQLPostgreSQL
PyTorchPyTorch
scikit-learn
LightGBM
Whisper
CrewAI
KotlinKotlin
MLOps
pgvector
Node.jsNode.js
ASP.NET Core
MongoDBMongoDB
MCP
GoGo
C#C#
Pydantic
Kaggle
Hugging Face
Independent project creation

Projects

WhatsApp MCP Server

personal
Built a personal high-value project creating an MCP server for WhatsApp to configure the backend and create functions for tool usage.
Gowhatmeow

Vessel

personal
Developed an API-first platform in .NET for inquiring about water tankers in high-demand locations like Karachi and Islamabad.
C#.NET

PSL Dataset and SignSpeak (Final Year Project)

Lead Developeracademic
Built one of the largest public word-level PSL landmark datasets (70+ classes, 7K sequences), as measured by 400+ downloads across Kaggle and Hugging Face, by designing the video capture and MediaPipe landmark extraction pipeline and publishing it for reproducibility. Trained a TensorFlow LSTM sign classifier to 92% accuracy on the dataset. Built a Kotlin Android app and FastAPI backend that translate camera-recorded signs to text, with an LLM turning detected words into full sentences.
TensorFlowPyTorchscikit-learnMediaPipeKotlinFastAPILLMMobile app

Education

B.S. Software Engineering

COMSATS University, AbbottabadSoftware EngineeringGPA 3.992022 – 2026

Thesis. PSL Dataset and SignSpeak

Societies & activities

Gold MedalistBatch TopperMerit scholarship for 7 semestersPM Laptop Scheme awardee1st Place, Mobile App Development Competition

Recognition

Awards

Gold Medalist

COMSATS University2026

Batch Topper

Merit Scholarship

COMSATS University

Awarded for 7 semesters

PM Laptop Scheme Awardee

Government of Pakistan

1st Place, Mobile App Development Competition

COMSATS University

Patents & publications

PSL Landmark Dataset

publication2025

Public word-level Pakistan Sign Language landmark dataset published on Kaggle and Hugging Face.

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

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

More people on Kanvis

Browse more peopleCreate your own Kanvis