Selected Work

Projects

Four production systems. Each one ships to real traffic.

Featured

Conversational Cross-Channel Booking Agent

A cross-channel booking agent built on LangGraph with stateful conversation context. Integrates WhatsApp, Instagram, Gmail, and Outlook into a single inbox. Orchestrates calendar syncing across Google and Microsoft calendars via OAuth flows. Sustains 99%+ uptime, cutting vendor scheduling overhead by roughly 1.5 hours daily and saving up to 8.5 hours per week.

LangGraphFastAPIDockerGCPRedisCeleryOAuth2
Case Study

Vision-RAG Mining Intelligence System

End-to-end RAG pipeline using Gemini 2.5 Flash multimodal capabilities to parse 300 DPI scanned mining reports and reconstruct tables where traditional OCR failed. FAISS dense embeddings combined with BM25 sparse search via Reciprocal Rank Fusion -93%+ accuracy on a 300+ document corpus with sub-3s latency.

Gemini 2.5 FlashFAISSBM25LangChain

Early Retirement -AI-Driven NPC Game

NPCs that respond to what you actually said instead of cycling through a dialogue tree. SmolLM2-360M, fine-tuned with QLoRA on 20,182 synthetic dialogues, running locally at ~17 MB. Full voice loop -speech in, model, speech out -at roughly 3 seconds end to end.

PyTorchQLoRASmolLM2Deepgram

Multilingual Sentiment Analysis Pipeline

Sentiment classification across English, Arabic -Modern Standard and Saudi dialect -and Urdu. Custom attention heads on AraBERT and RoBERTa-large recover morphology a general multilingual model misses. Trained on 120K+ labeled reviews, ablated against standard fine-tuning to isolate the gain. 96% F1 across all three, 14 points above off-the-shelf on Saudi dialect Arabic, serving live inference under 1s per review.

PyTorchAraBERTRoBERTaTransformers

Enterprise Agentic RAG with Social Connectors

A company-wide knowledge assistant that answers from internal documents, live Slack threads, Jira tickets, and Gmail records behind one query interface. WSO2 Integrator normalizes the connectors, the WSO2 AI Agent picks the tool at query time, and the vector store is partitioned by department so retrieval never crosses an access boundary. Sub-2s responses at 93%+ accuracy.

WSO2 IntegratorWSO2 AI AgentLangChainMilvus