N8N ARCHITECTUREVOICE AI AGENTSLLM ORCHESTRATIONPYTHON AUTOMATIONRAG SYSTEMSAPI INTEGRATIONN8N ARCHITECTUREVOICE AI AGENTSLLM ORCHESTRATIONPYTHON AUTOMATIONRAG SYSTEMSAPI INTEGRATION
0
Success Executions
0
Reliability Rate (%)
Agentic AI
Core Specialization
Profile
Beyond Basic Automation.
I build systems that don't just work—they scale. My focus is on creating autonomous logic that handles complex business processes with zero downtime.
From automating thousands of e-commerce product entries to building human-like voice responders, I bridge the gap between AI potential and production reality.
Selected Work
Featured Works
Enterprise / n8n
n8n E-Commerce Automation
Automated product title & description generation for a major retailer. 1,100+ production runs with zero errors.
AI / Computer Vision
FAISS Visual Search Engine
High-scale product search using local ONNX SigLIP embeddings and a FAISS index. Integrates MongoDB and AWS S3 with intelligent OCR fallback.
Voice AI
Autonomous Voice Agent
AI agent that handles live bookings, availability, and scheduling with real-time calendar synchronization.
Lead Generation
Hyper-Personalized Outreach
Scraping, website analysis, and personalized email generation workflow with LLM integration.
RAG / ML
Custom AI Knowledge Base
Internal document assistant using vector search and large language models for instant business intelligence.
The AI Product Import and Image Enhancement Pipeline was custom engineered for the Communets E-commerce Admin Panel. By integrating n8n workflow automation, local large language models (Qwen2.5:7b via Ollama), headless scraping engines, and state-of-the-art computer vision models (ESRGAN & BiRefNet), the pipeline entirely automates the product catalog ingestion process.
Production Impact
Successfully processed and imported 1,200+ real products directly into the client's production database, reducing onboarding times by over 96%.
Operational Bottlenecks (Before)
SEO Optimization & Copywriting: Manually writing descriptions and bullet points took roughly 10 minutes per SKU.
Specs Harvesting: Manually searching Google and manufacturer websites for specifications sheets.
Image Prep: Removing background clutter and resizing supplier photos manually using Photoshop.
Automated Solution (After)
Instant SEO Enrichment: Agentic web scraping and local LLMs write professional content in 20 seconds.
Self-Hosted Image pipeline: Automated ESRGAN upscaling and BiRefNet foreground separation.
AWS S3 Sync: Automatic conversion to transparent, optimized WebP format with secure cloud upload.
System Architecture
The system coordinates a concurrent data enrichment flow and an image processing flow managed by an Express.js background queue. Click on the boxes below to inspect the sub-systems.
Hover/Click a Component
Select any stage of the pipeline in the flowchart above to inspect technical specifics, tools, and protocols.
Performance Metrics & Business Impact
The transition from manual copywriting and Photoshop cropping to the autonomous AI pipeline delivered a step-change in scaling capabilities and operations efficiency.
0%
Time Reduction
0
Imported Products
0%
Execution Success
0
API Overheads ($)
Onboarding Phase
Manual Operations (Before)
Automated Pipeline (After)
Improvement
Product Copywriting
10 mins (manual SEO research)
8 secs (Ollama Qwen2.5 generation)
98% Faster
Image Editing & BG Cut
5 mins (Photoshop clipping path)
6 secs (Local BiRefNet model)
98% Faster
Format & Optimization
2 mins (Save for web, upload)
2 secs (Sharp rendering & S3 API)
90% Faster
Average Total/SKU
~17 Minutes
< 20 Seconds
96.6% Improvement
Project Overview
This case study covers the architecture behind a **High-Performance Hybrid Autocomplete and Multi-Tiered Visual Search Engine**. Built to support high-volume catalog discovery, it coordinates instant prefix/substring alphanumeric matching and a visual search pipeline that falls back to OCR and multimodal tagging in the event of visual matching misses.
Discovery Latency
The primary vector search path returns matching catalog elements in under 15 milliseconds by loading vectorized features directly into a local in-memory FAISS engine.
Engineering Challenges
Image Search Overhead: Processing heavy vision embeddings dynamically on the main Node API thread causes connection bottlenecks.
Heuristic Misses: Traditional vision searches return irrelevant items if the query image has a messy background or poor illumination.
SKU Autocomplete: Inability of standard database text indexes to match sub-words (e.g. searching "A" to find "AB-102").
Engine Architecture
FastAPI Microservice: Heavy mathematical computations (ONNX SigLIP) decoupled from the Express gateway.
Dual-Route Querying: RegEx string matches for short queries and SKU prefixes bypass heavy search operations.
Intelligent Autocomplete & SKU Routing
Alphanumeric SKU checks are routed directly to instant scan queries, falling soft to weighted compound database scans, and degrading to title substring regex fallback if no direct matches exist. Hover/click nodes to inspect.
Multi-Tier Discovery Pipeline
The visual search engine operates a multi-tiered fallback pipeline. The primary embedding match runs on local ONNX SigLIP & FAISS, failing soft to Gemini Vision tagging and OCR text extraction if similarity limits are not satisfied.
Hover/Click a Component
Select any level of the search engine above to inspect embeddings protocols, libraries, and latency metrics.
Performance Metrics & Business Impact
Through offloading calculations to python and building a multi-tier model cascade, the discovery latency remained in standard limits even when vector index similarity was not matched.