AI platforms, multi-agent systems, ML solutions, and full-stack applications.
Built a full SaaS platform enabling businesses to embed customizable AI assistants on their websites. Features a RAG pipeline using pgvector for cosine similarity search with sentence-transformers for embeddings, supporting PDF/text/URL ingestion. Includes real-time SSE chat streaming, an embeddable React/TypeScript widget bundled as a single IIFE script, intelligent lead capture with keyword-based intent detection and webhook integration, an admin dashboard with analytics and CSV export, a WordPress plugin for one-click integration, and security hardening including SSRF protection, prompt injection sanitization, and rate limiting.
Designed and developed a full-stack multi-tenant invoicing and ERP platform with an integrated AI agent system (orchestrator/planner/executor pattern) for natural language business queries. Supports full document lifecycle management for Invoices, Quotations, Proforma Invoices, Credit Notes, Purchase Orders, and Recurring Invoices. Includes financial reporting (revenue, P&L, aging, balance sheets, ledger views), server-side PDF generation with customizable templates, multi-currency support with exchange rate integration, vendor & purchase order management, customer directory with statement generation, and trilingual support (English, Arabic, Urdu).
Architected and implemented a distributed multi-agent system leveraging CrewAI to automate customer service workflows, enabling parallel processing of support tickets and intelligent task delegation. Engineered specialized agents for ticket classification, knowledge retrieval, and response generation, resulting in a 40% reduction in average response time and improved customer satisfaction metrics.
Engineered an end-to-end automation pipeline utilizing computer vision and large language models to process custom warehouse documents. The system extracts and transforms unstructured data from PDFs, Word Documents, and other formats into structured JSON, enabling seamless data integration with 90% accuracy and resulting in a 60% improvement in operational efficiency.
Architected and implemented an advanced AI-powered ERP solution leveraging Frappe ERPNext framework with WhatsApp integration, enabling seamless user interaction through Natural Language Processing. The system streamlined business processes by allowing users to perform complex ERP functions through conversational interfaces.
Built an AI-powered RESTful API for validating and correcting customs consolidation data used in customs declarations. Designed a multi-phase validation engine: Phase I corrects product tags for consistency, grammar, and customs compliance; Phase II flags HS code misclassifications. Integrated OpenAI GPT with domain-specific prompt engineering for customs classification. Implemented concurrent async processing with semaphore-based rate limiting to handle multiple declarations in parallel. Includes JSON payload validation, file-based validation, and validate-and-save with automatic output generation.
Built a backend microservice for an accounting/finance platform that automates financial data extraction, investment report parsing, transaction categorization, and travel document processing. Integrated Plaid API for real-time bank transaction fetching with automatic personal vs. business classification. Developed an AI-powered investment report parser using GPT-4o Vision API supporting Morgan Stanley, Robinhood, MidFirst Trust, and Krypton Fund Services. Built a two-stage AI pipeline for travel document classification and structured data extraction. Includes Excel report parser, Dropbox integration, and containerized deployment with Gunicorn.
Built an MCP (Model Context Protocol) server enabling AI assistants like Claude to programmatically search, browse, and extract product data from McMaster-Carr's industrial supply catalog. Implemented browser automation with Puppeteer to overcome API limitations. Designed 7 MCP tools: product search, product details, category browsing, dynamic filter discovery, filter application, product variants, and category navigation. Built a dynamic filter discovery system extracting category-specific filters from the DOM. Engineered dual-format output (Markdown & JSON) with Zod-validated input schemas. Supports Stdio for Claude Desktop and HTTP/Express for remote access.
Built a production-grade AI-powered data extraction system that processes shipping labels, invoices, and logistics documents (PDFs, images, HEIC) to extract structured information. Designed a multi-stage pipeline: image preprocessing, OCR/vision analysis, structured JSON extraction, and regex-based post-processing. Integrated GPT-5 vision models for text extraction, barcode/QR code detection (CODE128, QRCODE) for tracking numbers across 8+ carriers (UPS, USPS, FedEx, DHL, Amazon), fuzzy string matching for merchant validation, and async batch processing with multi-file support. Deployed on AWS with CI/CD via GitLab and CodeDeploy.
Built an end-to-end music analytics platform comprising a multi-stage ETL pipeline and an API layer. The data pipeline transfers Chartmetric analytics from AWS S3 to Google BigQuery, processing 25+ entity types with incremental processing, dual-layer deduplication (set-based file tracking + BigQuery DISTINCT queries), and automated VM lifecycle management via Cloud Functions and Cloud Run Jobs. Covers Spotify, Deezer, iTunes playlist stats, artist audience metrics, TikTok/Instagram/YouTube analytics, and track relationships. Optimized with GLACIER file skipping, BigQuery table clustering, and SNAPPY-compressed Parquet handling. The API layer includes a FastAPI ingestion service converting Spotify Artist IDs to Chartmetric IDs across 5 BigQuery tables, and a Flask analytics API exposing 50+ endpoints with Redis caching, JWT authentication, batch processing for up to 50 concurrent artist ingestions, and structured logging with request tracking.
Built a production-grade AI pipeline that extracts structured data from multi-format invoices and classifies products with HS tariff codes using multi-model AI. Designed a 9-phase extraction pipeline with cascading model fallback (Claude → GPT-5 → GPT-5-mini → ChatGPT-4o) ensuring 99%+ uptime. Built a dual-approach HS code classification engine combining FAISS + sentence-transformers embedding search with LLM-based classification. Supports PDF, images, DOC/DOCX, MSG/EML emails. Implemented concurrency controls with semaphore-based thread limiting (20 file workers, 6 prediction workers), comprehensive data validation with subtotal verification and currency conversion, and token usage tracking with cost monitoring.
Built an AI-powered REST API that automatically categorizes bank transactions into a business's custom accounting categories using OpenAI GPT. Implemented concurrent batch processing with ThreadPoolExecutor (20 workers), processing up to 10 transactions per batch with semaphore + lock-based throttling to stay within API quotas. Built robust retry logic with exponential backoff and jitter. Engineered a multi-strategy category matching system with exact match, partial/fuzzy match, and graceful fallback. Integrated real-time cost tracking per API request logging token usage and per-transaction metrics. Containerized with Docker and Docker Compose.
Architected and implemented a distributed multi-agent system leveraging CrewAI framework to autonomously research, synthesize, and generate comprehensive articles, streamlining content creation workflows with specialized agents for research, writing, and editing.
Designed and implemented a high-accuracy prediction model using LightGBM, achieving 85% accuracy in forecasting movie revenues through feature engineering and advanced regression techniques.
Engineered a context-aware meditation guidance system implementing Retrieval Augmented Generation (RAG) architecture with LangChain, delivering personalized mindfulness experiences based on user context and preferences.
Architected a real-time fraud detection platform incorporating machine learning algorithms to identify and flag suspicious financial activities, enhancing transaction security protocols with automated anomaly detection and alerting.
Spearheaded the development of an AI-driven inventory optimization solution, implementing predictive analytics that reduced holding costs by 20% through improved demand forecasting and automated reorder point calculations.