Haris Umer

AI/ML Engineer specializing in LLM Systems, RAG, and Production Backend Architecture.

Based in: Lahore, PakistanMost recently: AI/ML Engineer, CoalDevExperience: 2 years

Experienced AI/ML Engineer with a proven track record of building high-throughput trading engines and production-grade RAG systems. He specializes in integrating advanced computer vision and LLM workflows into scalable Python backends using tools like Django, Celery, and AWS.

Experience

AI/ML Engineer

CoalDevLahore, PakistanworkJul 2025 – Sep 2026

Summary

Re-architected a real-time trading engine on Redis pub/sub and Celery to process ~6,000 market signals per second, raising throughput by an estimated 70-80%. Designed a multi-manager architecture that keeps live trading state in sync across accounts and strategies, and fixed PostgreSQL and Redis connection bottlenecks under peak load. Delivered Sortsy (sortsy.ai) features across 3 surfaces: Django REST API, Next.js web app and PySide6 desktop agent, backed by Celery workers and WebSockets. Shipped a serverless vision pipeline on AWS Lambda and Amazon Bedrock that captions files and indexes them in Qdrant for semantic file search and retrieval. Deployed production n8n agent workflows with Anthropic Claude, Gemini embeddings and pgvector RAG.

Results

Raised throughput by an estimated 70-80% for a real-time trading engine.
Processed ~6,000 market signals per second.

RedisCeleryPostgreSQLDjango REST APINext.jsPySide6WebSocketsAWS LambdaAmazon BedrockQdrantn8nAnthropic ClaudeGeminipgvectorRAGEC2CI/CD

AI/ML Engineer

Nexpred SolutionsLahore, PakistanworkApr 2025 – Jun 2025

Summary

Developed real-time multi-camera passenger analytics for 2 airport zones using YOLO, ByteTrack and homography, logging trajectories to JSON for queue-density analysis. Built an LLM banking assistant with Gemini tool calling for 3 workflows: authentication, disputes and loans. Optimized aircraft parking and departure scheduling with a reinforcement learning agent in a 3D hangar simulation.

What I did

Tracked people to measure time spent in queues, distance covered, and duration at specific counters.
Tested and compared Grounding DINO and self-trained YOLO models for passenger identification and tracking.
Built a 3D aircraft hangar environment and simulation for reinforcement learning experiments.
Developed a simulation framework to test sequential scheduling decisions based on environment states.

YOLOByteTrackHomographyJSONLLMGeminiReinforcement LearningGrounding DINO

AI/ML Engineer

BitPix SoftLahore, PakistanworkJun 2024 – Sep 2024

Summary

Engineered hierarchical chunking and agentic RAG for a legal-advisory assistant, one of 3 RAG chatbots built in 4 months, to improve answer precision. Integrated 2 speech pipelines (speech-to-text and text-to-speech) to power voice-enabled chatbots. Automated synthetic training-data generation with LLMs to expand chatbot response coverage.

What I did

Implemented hierarchical chunking and agentic RAG to handle large legal clauses within small context windows.

Results

Built 3 RAG chatbots in 4 months.

RAGSpeech-to-textText-to-speechLLMsGPT-3

Skills

Technical

Python

RAG

AI Agents

LLMs

LangGraph

OpenAI

Anthropic Claude

Google Gemini

YOLO

Embeddings

Semantic Search

Redis

Django REST Framework

Celery

REST APIs

Amazon Bedrock

WebSockets

Event-Driven Architecture

Tool Calling

AWS Lambda

Prompt Engineering

LangChain

Qdrant

pgvector

Transformers

Git

PostgreSQL

Fine-Tuning

ByteTrack

Stable Diffusion

FastAPI

OpenCV

Hugging Face

JavaScript

Docker

Reinforcement Learning

PyTorch

SQL

Flask

TypeScript

LlamaIndex

TensorFlow

CI/CD

C++

Java

n8n

Railway

GPT-3

Grounding DINO

Projects

Visible Electrical Defect Detection

Computer Vision Engineerpersonal

Trained YOLOv8n detectors on 7,000+ labeled images to find open, short and damage defects in PCBs and cables. Scored 0.888 mAP50 on outdoor cable damage and 0.831 mAP50 on a 4-class PCB and cable model. Benchmarked a CLAHE/sharpen enhancement pass on 484 test images, raising precision from 0.81 to 0.86.

YOLOv8nComputer VisionCLAHE

Aircraft Parts Quote Automation

Lead Architectprofessional

Architected an LLM pipeline that scans up to 50 Gmail threads per part into Pydantic-validated structured outputs. Ranked the top 5 unique suppliers per part across 5 Google Sheets tabs, replacing manual email review. Validated on real part numbers with 50-96 supplier emails each, processing parts in batches of 10.

LLMPydanticGmail APIGoogle Sheets

Artikon - Agentic Comic Generator

Final Year Project Studentacademic

Fine-tuned SDXL with IP-Adapter and ControlNet on a custom image dataset for consistent comic characters. Orchestrated a 7-node LangGraph character subgraph for prompts, masks, generation and compositing. Implemented embedding search that maps each scene to up to 10 LoRA style tags, turning scripts into comic pages.

SDXLIP-AdapterControlNetLangGraphLoRAEmbedding Search

n8n Automation Workflows

personal

Created self-learning workflows in n8n connected to PostgreSQL to fetch and update specific data.

n8nPostgreSQL

Web Crawler

personal

Built a web crawler for self-learning purposes.

Education

Bachelor of Science

FAST-NUCESComputer ScienceAug 2021 – Jun 2025

Thesis. Artikon - Agentic Comic Generator

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

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

More people on Kanvis

Browse more people Create your own Kanvis