You will get a real-time object detection and tracking system
Project details
Most tracking demos look fine on clean sample footage and fall apart on real cameras. IDs switch when people cross, objects vanish behind obstacles, and frame rates collapse on actual hardware.
Background
I co-authored a practical evaluation framework for real-time multi-object tracking published in IEEE Access, and I build production vision systems at an AI company in South Korea, including retail shelf monitoring and a face recognition platform running on live camera networks.
How it works
I choose a detector and tracker pairing suited to your footage rather than defaulting to one stack, tune it on your own video, and report HOTA, ID switches, and FPS on your data. For edge devices, I convert to ONNX and TensorRT with quantization and benchmark the speed and accuracy tradeoff on your actual hardware.
Typical uses
People counting and dwell time, vehicle and traffic analysis, retail product tracking, workflow and safety monitoring, and multi-camera setups where identity must persist across views.
Stack
PyTorch, OpenCV, YOLOv8 to v11, YOLOX, ByteTrack, BoTSORT, DeepOCSORT, StrongSORT, TorchReID, ONNX, TensorRT, NVIDIA Jetson.
Background
I co-authored a practical evaluation framework for real-time multi-object tracking published in IEEE Access, and I build production vision systems at an AI company in South Korea, including retail shelf monitoring and a face recognition platform running on live camera networks.
How it works
I choose a detector and tracker pairing suited to your footage rather than defaulting to one stack, tune it on your own video, and report HOTA, ID switches, and FPS on your data. For edge devices, I convert to ONNX and TensorRT with quantization and benchmark the speed and accuracy tradeoff on your actual hardware.
Typical uses
People counting and dwell time, vehicle and traffic analysis, retail product tracking, workflow and safety monitoring, and multi-camera setups where identity must persist across views.
Stack
PyTorch, OpenCV, YOLOv8 to v11, YOLOX, ByteTrack, BoTSORT, DeepOCSORT, StrongSORT, TorchReID, ONNX, TensorRT, NVIDIA Jetson.
Machine Learning Tools
Deeplearning4j, Keras, NumPy, OpenCV, Python, PyTorch, scikit-learn, TensorFlowWhat's included
| Service Tiers |
Starter
$50
|
Standard
$250
|
Advanced
$1,000
|
|---|---|---|---|
| Delivery Time | 3 days | 5 days | 7 days |
Number of Revisions | 1 | 3 | Unlimited |
Number of Model Variations | 2 | 3 | 5 |
Number of Scenarios | 1 | ||
Model Validation/Testing | |||
Model Documentation | |||
Data Source Connectivity | |||
Source Code |
Optional add-ons
You can add these on the next page.
Fast Delivery
+$50 - $100
Additional Revision
+$30
Additional Model Variation
(+ 1 Day)
+$30
Additional Scenario
(+ 2 Days)
+$20
Production Relaease
(+ 3 Days)
+$300Frequently asked questions
About Mirsaid
AI Developer | LLMs, RAG Chatbots, AI Agents | Computer Vision, YOLO
Incheon, South Korea - 7:40 am local time
I build systems that work with real users, real business data, and real production constraints - not just demos or notebook experiments.
🚀 One project I’m especially proud of is an agentic legal AI assistant serving 7,000+ users and 500+ paying subscribers, handling 15,000+ messages per month across hundreds of thousands of documents in 4 languages.
I led the AI architecture and RAG development, including ingestion, retrieval, reranking, prompting, and evaluation. We improved answer accuracy from 38% to 51% and outperformed competing legal AI assistants by 17% across 22+ evaluation categories.
I’ve also shipped:
🎙️ A native voice AI Agent with sub-3-second responses, running 24/7 across 6+ industries
🤖 A multi-agent LLM system that processed 65.5M product descriptions
👁️ Computer Vision systems for object detection, tracking, face recognition, retail monitoring, and anomaly detection
🤖 AI CHATBOTS & RAG
I build custom assistants using OpenAI GPT, Claude, Gemini, Llama, DeepSeek, and open-source models.
Use cases include customer support, sales, lead qualification, internal knowledge assistants, document chatbots, AI copilots, and multilingual assistants.
Your AI can securely work with PDFs, websites, Notion, Confluence, SharePoint, SQL databases, APIs, and internal knowledge bases.
I can add citations, memory, human handoff, guardrails, structured outputs, streaming, analytics, and monitoring.
🔎 ADVANCED RAG
I work with semantic search, hybrid search, reranking, metadata filtering, query rewriting, multi-stage retrieval, GraphRAG, and knowledge graphs.
Vector/search stack: Pinecone, Qdrant, Milvus/Zilliz, pgvector, Elasticsearch, ChromaDB, Neo4j.
I also use RAGAS, Langfuse, retrieval evaluation, hallucination tracking, and prompt testing so you can measure whether the system is actually improving.
🔌 AI AGENTS & AUTOMATION
I integrate LLMs with CRM, ERP, Zendesk, Intercom, HubSpot, Stripe, Google Workspace, Airtable, databases, custom APIs, and webhooks.
I build function/tool calling, MCP servers, structured workflows, agent routing, retries, safe fallbacks, and multi-agent systems using LangGraph, LangChain, CrewAI, and custom orchestration.
🎙️ VOICE AI
I build low-latency voice agents for customer support, reception, lead qualification, sales, and appointment workflows using Whisper, ElevenLabs, Twilio, streaming APIs, WebSockets, and custom STT/TTS pipelines.
👁️ COMPUTER VISION
I build real-time object detection, multi-object tracking, ReID, face recognition, retail monitoring, smoke/anomaly detection, and Edge AI systems.
I also co-authored a peer-reviewed IEEE Access paper introducing LITE, a lightweight feature extraction method for faster ReID-based tracking.
Stack: PyTorch, OpenCV, YOLOv8–YOLO11, ByteTrack, BoTSORT, DeepOCSORT, StrongSORT, InsightFace, ONNX, TensorRT, NVIDIA Jetson.
⚙️ PRODUCTION & BACKEND
I also handle the engineering needed to make AI reliable in production: Python, FastAPI, async I/O, streaming, PostgreSQL, Celery, Prefect, Docker, AWS, GCP, Azure, monitoring, caching, model routing, and cost optimization.
💬 WHAT IT’S LIKE WORKING WITH ME
I communicate clearly, scope honestly, and focus on practical solutions.
If a simpler approach can deliver 90% of the value for one-third of the cost, I’ll tell you before you spend the money.
Have an idea, a half-built AI product, or a messy problem you think AI could solve?
Send me a message and tell me what you’re trying to build. I’ll help you figure out the most practical way to turn it into a reliable production system.
🚀 Let’s build something people actually use.
Steps for completing your project
After purchasing the project, send requirements so Mirsaid can start the project.
Delivery time starts when Mirsaid receives requirements from you.
Mirsaid works on your project following the steps below.
Revisions may occur after the delivery date.
Footage review and feasibility check
I review your sample video and requirements, assess camera angles, resolution, and object density, and confirm what accuracy is realistically achievable. You receive a short plan stating what the system will detect and what its known limitations.
Detector selection and tuning
I test candidate detection models on your footage and tune confidence thresholds, input resolution, and class filtering for your specific conditions. This determines the ceiling on tracking quality, so it is done before any tracking work begins.