Muhammad Tahir Ali

AI Engineer specializing in high-throughput OCR pipelines and agentic AI systems.

Based in: Islamabad, PakistanCurrently: AI Engineer, LimeoXExperience: 3 years

Hire me

Muhammad Tahir Ali is an experienced AI Engineer with a strong background in building large-scale OCR analytics, agentic voice systems, and medical image synthesis using GANs. He excels at deploying complex machine learning models into production environments using tools like LangGraph, FastAPI, and Docker.

Experience

AI Engineer

LimeoXSargodha, PakistanworkNov 2025 – Present

Summary

Built a high-throughput OCR pipeline processing 35,000+ images daily, turning noisy screenshots into structured, searchable document timelines using Fara-7B on RTX 3090/5090 GPUs. Built a LangGraph-based AI email replier for recruiting, expanding intent classification from 15 to 29 categories and refining prompts. Developed an NLP keyword-extraction pipeline for a content platform, benchmarked against Surfer SEO, fixing classifier, scorer, and extractor issues.

What I did

  • Built the keyword pipeline, the Semantic Content platform, and the OCR pipeline.
  • Deployed the LangGraph HR agent as Dockerized FastAPI services with PostgreSQL audit logging.
  • Implemented a LangGraph HR email agent using Pydantic models to enforce structured outputs for intent classification and field extraction.
  • Served models using vLLM with continuous batching and PagedAttention for high-throughput processing.
  • Utilized NVIDIA Triton Inference Server for concurrent request handling and dynamic batching under tight VRAM constraints.
  • Used Hugging Face Transformers for model prototyping and debugging.

Results

  • Evaluated 8-bit versus 16-bit quantization for Fara-7B on OCR tasks, opting for 16-bit to maintain output quality.
  • Processed 35,000+ images daily in OCR pipeline.
  • Expanded intent classification from 15 to 29 categories for AI email replier.
OCRFara-7BRTX 3090RTX 5090LangGraphNLPSurfer SEODockerFastAPIPostgreSQLPydanticJSON modeStructured outputvLLMNVIDIA Triton Inference ServerHugging Face TransformersQuantization

AI Engineer

Aurelion.aiRemote, TorontoworkJul 2025 – Nov 2025

Summary

Built Dialer Desk, an agentic AI voice-calling system that handles phone conversations autonomously, in the vein of Retell AI. Developed an agentic AI code reviewer that analyzes pull requests and posts actionable review comments, in the vein of CodeRabbit.

What I did

  • Tracked end-to-end response latency per turn at p50 and p95 for the voice-calling agent.
  • Monitored call completion rate, tool-call failure rate, and the handling of interruptions and barge-ins.
  • Tracked acceptance rate of review comments, false-positive rates on flagged issues, latency per PR, and token cost per review for the PR reviewer.
  • Tightened prompts and added filtering for the PR reviewer to focus on high-confidence issues when comment acceptance rates were low.
Agentic AIVoice AIRetell AICodeRabbit

Associate AI Researcher

BIOMISA LabIslamabad, PakistanworkAug 2024 – May 2025

Summary

Achieved SSIM 0.9598 and PSNR 36.23 dB using Pix2Pix cGAN on 37 paired CT-MRI scans. Deployed on TI DSP Kit. Trained 48M+ parameter UNet generator for high-detail medical image synthesis.

What I did

  • Deployed on TI DSP Kit.
Pix2Pix cGANTI DSP KitUNetMedical Image Synthesis

Skills

Technical

PythonPython
Computer Vision
NLP
OCR Pipelines
FastAPIFastAPI
LangChain
LangGraph
OpenCVOpenCV
Scikit-learn
GitGit
GANs
RAG
PyTorchPyTorch
TensorFlowTensorFlow
Next.jsNext.js
SQLSQL
PostgreSQLPostgreSQL
DockerDocker
TypeScriptTypeScript
AWS
MLflow
CI/CD
NVIDIA Triton Inference Server
GitHub ActionsGitHub Actions
Quantization
LiveKit Agents
AWQ
Pydantic
JSON mode
Structured output
vLLM
Hugging Face Transformers
k-means classification

Projects

Large-Scale Visual Intelligence and OCR Analytics Platform

Lead Developerprofessional
Built a high-throughput OCR-based image-to-data pipeline processing 35,000+ images daily, extracting structured metadata and consolidating noisy screenshot text into searchable document-level timelines. Used Microsoft/Fara-7B model for this on RTX 3090 and RTX 5090 GPUs.
OCRMicrosoft/Fara-7BRTX 3090RTX 5090

CT to MRI Synthesis via GAN

Researcheracademic
Achieved ~0.96 SSIM in MRI synthesis from CT scans using Pix2Pix GAN; deployed on a TMS320 DSP kit in C for embedded real-time inference.
Pix2Pix GANTMS320 DSP kitC

Multi-Tenant Voice AI Dialer

Full Stack AI Developerprofessional
Built a multi-tenant voice AI calling platform with Next.js, TypeScript, Python, LiveKit Agents, and PostgreSQL, enabling real-time outbound calls, node-based conversation workflows, RAG-powered knowledge lookup, post-call transcript and analytics pipelines.
Next.jsTypeScriptPythonLiveKit AgentsPostgreSQLRAG

Semantic Content Intelligence Platform

Engineerprofessional
Engineered an end-to-end NLP pipeline using GLiNER, BGE embeddings, and reranking models to convert large-scale web content into ranked entities, topic clusters, and structured decision-ready signals.
NLPGLiNERBGE embeddingsReranking models

Kubernetes ML deployment project

personal
A project completed at NUST involving the deployment of machine learning models using Kubernetes.
Kubernetes

Education

Bachelor of Electrical Engineering

National University of Science & TechnologyElectrical EngineeringJun 2025
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