Engineering Archive.

For global logistics firms, the manual processing and compliance verification of cross-border documentation (Bills of Lading, Commercial Invoices, Packing Lists) is a multi-day bottleneck prone to costly human error. I architected and deployed the Agentic Document Compliance Pipeline, a state-of-the-art, 6-agent orchestration system powered by LangGraph. By fine-tuning LLaMA-3 with QLoRA and routing verifications through a custom Model Context Protocol (MCP) layer, the pipeline cut manual document-processing time by ~92%, raising extraction accuracy to 96% F1 while drastically reducing hallucination and false-approval rates.

Deploying heavy Deep Learning models in production often results in unacceptable latency and bloated cloud compute costs if architectural patterns are not modernized. The Azure GPU-Backed Streaming Inference Migration represents a masterclass in MLOps engineering: the end-to-end migration of a legacy, on-premise CPU batch-inference system to a high-throughput, GPU-backed streaming microservice on Azure Kubernetes Service (AKS). By implementing Server-Sent Events (SSE) via FastAPI, dynamic tensor batching, and automated Blue-Green failover protocols, this migration cut inference latency by 92% and compute costs by 64%.

Deploying generative AI in customer support requires a delicate balance between speed, cost, and factual accuracy. BiliRAG is a production-grade, highly concurrent conversational assistant designed to serve bilingual (Urdu & English) customer support and logistics queries. By heavily optimizing the vector retrieval engine with FAISS HNSW indexing and applying ONNX INT8 weight quantization, the system achieves sub-60ms retrieval latencies and a 75% memory footprint reduction. Most importantly, it implements a rigid confidence-based routing layer that nearly eliminates LLM hallucinations, securing the trust of enterprise clients.

For founders, engineers, and high-output professionals, information overload is a critical bottleneck. Coeus is a robust Personal Knowledge Management (PKM) system built entirely on local, plain-text Markdown within Obsidian. Designed around Tiago Forte's PARA method (Projects, Areas, Resources, Archive), Coeus acts as a highly structured "Second Brain," allowing for frictionless capture, systematic organization, and rapid retrieval of complex technical, business, and personal data.

For independent consultants and agencies, the speed and quality of proposal generation directly dictate lead conversion rates. I engineered Ezra Bid Assistant, a secure, private Chrome extension designed to extract context from live freelance platforms and instantly generate highly personalized, high-converting proposals using the Google Gemini LLM. By decoupling the AI logic into a private Next.js backend, the system guarantees API key security while offering an intuitive, in-browser interface that accelerates the bidding workflow by an estimated 80%.

Ezra Global is a premium digital agency that blends "Brand Alchemy" with "Creative Engineering." To reflect their high-end market positioning, they required a web presence that was not just fast, but visually arresting and immaculately optimized for search engines. I architected and delivered a state-of-the-art landing platform using the latest Next.js 16 and React 19 features, ensuring an unparalleled user experience characterized by fluid animations, perfect accessibility, and zero-latency lead generation pipelines.

For enterprise customer support systems, a low-accuracy FAQ search forces users to open manual support tickets, driving up operational costs. The FAQ Search Reranking & Active Learning Prototype is a production-grade Search and Reranking prototype designed to rescue legacy keyword-based search systems. By introducing a multi-stage TF-IDF feature reranking pipeline and a Semi-Automated Active Learning loop, this architecture boosted FAQ retrieval accuracy by 18 percentage points and cut the manual data-labeling backlog from 3,000 to under 200 tickets.

Delivering sub-second, highly accurate responses to customer support queries across multiple languages requires robust vector search capabilities. The FAISS-Based FAQ Retrieval Engine is a production-grade, bilingual (English & Urdu) semantic retrieval engine that eliminates the need for expensive, high-latency translation layers. By combining Hugging Face Sentence-Transformers with a highly optimized FAISS HNSW (Hierarchical Navigable Small World) index, the system slashed retrieval latencies from 340ms to under 60ms while aggressively reducing LLM hallucinations through strict similarity containment gating.

As Physical AI (robotics, autonomous systems) rapidly transitions to training via synthetic data and simulated world models, the industry faces a crisis of trust: labs self-report benchmarks, and synthetic environments often suffer from undetected physics inconsistencies. Haga is a complete, end-to-end Physical AI verification platform comprising Haga Core (a Python-based verification layer that stress-tests robot learning policies in MuJoCo) and Haga Web (a secure PNPM monorepo with dual Next.js surfaces for public research marketing and private investor diligence). Together, this architecture delivers independent, third-party validation for Physical AI startups while airgapping source code and gating sensitive telemetry.

For enterprise organizations constrained by strict data-residency requirements, passing sensitive data to external APIs (like OpenAI) is a non-starter. The On-Prem LLM Fine-Tuning & Serving Framework is a standardized platform for on-premise QLoRA fine-tuning and dynamic vLLM-based serving. By standardizing the fine-tuning pipelines and leveraging dynamic LoRA adapter-swapping, the framework cut model-to-production time from 6 weeks to 9 days, driving massive GPU memory savings (45%) and empowering product teams with self-service AI capabilities.

In financial technology, fraud detection is a constant race against time. Systems must score millions of transactions with extreme accuracy; a delayed response frustrates users, while a false decline damages merchant relationships. The Real-Time Fraud Scoring Pipeline is a high-throughput, low-latency streaming system that successfully scores 1.2M+ daily transactions at < 180ms p95 latency. By implementing a two-tier ensemble classifier and automated feature-drift monitoring, the pipeline reduced false-positive rates by 75% and slashed production incident recovery times from ~50 minutes to under 8 minutes.

For modern retail businesses like beauty salons, high-friction point-of-sale systems directly impact the bottom line. I engineered a robust, cloud-native POS system for Suraya Beauty Point Salon tailored specifically for multi-branch operations, real-time analytics, and seamless bilingual support (English/Arabic). Built on the bleeding edge of the JavaScript ecosystem (Next.js 16, Bun, Neon Serverless Postgres), the application provides enterprise-grade performance, scalable database architecture, and a modern, high-conversion user experience.

Customer support analytics in regions spanning multiple languages and scripts (e.g., Urdu, Roman Urdu, and English) often suffer from severe data fragmentation. The Bilingual Urdu-English Sentiment Classifier is a production-grade end-to-end Machine Learning pipeline designed to unify and automate sentiment classification across these diverse linguistic inputs. By implementing robust Unicode normalization, an Active Learning uncertainty engine, and a FastAPI-driven auto-tagger, the system achieved 89% accuracy on a 12,000-ticket dataset, reduced manual feedback-tagging effort by 70%, and crushed a 3,000-ticket backlog down to under 200 items.