Projects & Case Studies

Architecture breakdowns, evaluation methodology, and live repositories across Generative AI, Agentic AI, Machine Learning, and Full-Stack Systems.

Multi-document RAG architecture diagram with ingestion pipeline, vector database similarity search, and grounded LLM evaluation dashboard
Generative AI · Ladder7 R&D

Multi-Document RAG Research Assistant

Vector-grounded LLM assistant validated for relevance, faithfulness, consistency, and hallucination control using LLM-as-a-Judge, BLEU, and ROUGE.

Agentic AI automation workflow diagram showing web scraper agents, eligibility screening, automated reasoning module, and scheduling telemetry
Agentic AI · Wadi Al Ain Production

Agentic Grant-Monitoring System

Python and browser automation workflow tracking deadlines and assessing eligibility across 5+ funding sources with Django & SQL integrity validation.

Enterprise NLP analytics dashboard showing 57,391 review topic clusters and 82.94% accuracy emotion metrics
Data Science · First Quadrant Labs

Review Sentiment & Topic Analytics

57,391 cleaned software reviews analysed with 82.94% emotion classification baseline, CSV validation, duplicate detection, and PM prioritization evidence.

Interactive System Walkthrough

AI Architecture Blueprint Explorer

Ladder7 Nextstep Solutions LLP · R&D System

Grounded research workflow turning multi-document corpora into cited answers tested for faithfulness and hallucination resistance.

Stage 01: Document Chunking — Ingestion & Tokenization

Splits multi-document research files into semantically coherent passages with controlled overlap and metadata tracking before indexing.

Python Tokenization Document Parsing
Verification Check: Verified boundary preservation across multi-page source documents.

All Systems & Repositories (3D Showcase)

Interactive 3D Stack · Drag, Click Side Cards, or Use Arrow Keys

System Design Decisions & Trade-offs

Architecture Choices Across Shipped Builds
Decision 01 · Resume Extraction & ARC

Typed Pydantic Contracts Over Free-Form JSON Prompts

Prompting an LLM to "return valid JSON" breaks downstream parsers when nested fields drift. Pairing Groq low-latency inference with Instructor and strict Pydantic schemas enforces exact types, reading order, and structured retry correction before responses reach the FastAPI layer.

PyMuPDF · Groq API · Instructor · Pydantic
Decision 02 · Multi-Document RAG Assistant

Combined Rubric + N-Gram Eval Harness

Single-score evaluation misses subtle citation drift in multi-document RAG. Combining retrieved-context grounding and embedding similarity with BLEU, ROUGE, and LLM-as-a-Judge test scenarios captures both semantic faithfulness and exact factual overlap before Docker deployment.

Vector Search · LLM-as-a-Judge · BLEU · ROUGE
Decision 03 · Documint ERP

Database-Layer Multi-Tenant RBAC via PostgreSQL RLS

Instead of relying solely on frontend route guards between Admin and Employee portals, Documint ERP enforces row isolation inside Supabase PostgreSQL Row Level Security (RLS) and PLpgSQL policies so tenant boundaries hold regardless of client state.

Supabase · PostgreSQL RLS · TypeScript · Bun
Decision 04 · Review Intelligence & Churn ML

Statistical Feature Pruning & Duplicate-Safe Ingestion

Rather than feeding raw noisy tables directly into complex models, Telecom Churn combines Chi-square, RFE, and ANOVA F-tests before stacking classification, while First Quadrant Labs' review pipeline validates CSV schemas and deduplicates before persisting reusable model artifacts.

Scikit-Learn · Pandas · Feature Selection · CSV Validation