Opeyemi Momodu

Self-correcting RAG agent built as a LangGraph state machine (agent → retrieve → grade → rewrite/generate): the LLM decides whether retrieval is needed, a Pydantic structured-output judge gives a strict yes/no on chunk relevance, and the agent rewrites the query and retries when context is weak — replacing the brittle retrieve → stuff → generate pipeline. Hardened with a bounded rewrite budget and recursion limit to force a final answer instead of looping forever. Retrieval is exposed as a tool (WebBaseLoader → splitter → FAISS → ToolNode), with a provider-agnostic LLM/embeddings factory supporting fully-local, zero-key execution. Streamlit UI streams each node's trace live — tool-call decisions, retrieved chunks with sources, grader verdict, query rewrites, and final answer.

Production-grade document processing pipeline converting unstructured financial PDFs into validated structured data for regulated private banking. Compliance-aware processing with automated PII masking (Presidio), full audit logging, and human-in-the-loop review. Modular orchestrator with discrete steps for OCR (docTR/Tesseract), LLM extraction with structured JSON validation, risk scoring, and field-level confidence assessment — end-to-end traceability via Pydantic schema enforcement and a risk detection framework that flags anomalies for analyst review.

Production-grade multi-agent system coordinating specialized agents (retrieval, reasoning, validation, verification) for complex workflow automation. Features circuit breakers, exponential backoff, async scalable API layer, end-to-end observability with OpenTelemetry, and AI-specific QA — hallucination detection, reasoning fidelity scoring, and automated evaluation.

Production-grade AI system automating logistics documents (CMR, invoices, delivery notes). Pluggable LLM architecture (OpenAI / Ollama) for flexible provider switching and cost optimization. OCR + validation pipeline with high-accuracy extraction, human-in-the-loop review, and event-driven downstream integrations with retry logic.

AI-driven company analysis platform simulating a VC analyst workflow. Agentic scoring system evaluating startups across 5 dimensions using structured LLM outputs. Data ingestion pipelines converting unstructured data into structured insights, with configurable prompts turning business strategy into executable AI logic.

Serverless RAG system on AWS Lambda using LangChain, ChromaDB, and AWS Bedrock (Claude 3 Haiku + Amazon Titan embeddings). Automated document ingestion with deterministic chunk IDs for zero-downtime refresh. End-to-end CI/CD via GitHub Actions — pytest, Docker build check, then ECR push on merge.
NB: Private doc used as knowledge base in demo.

Retrieval-augmented chatbot with semantic vector search (HuggingFace Sentence Transformers), multi-domain support (game rules, platform guidance, Q&A), deployed to Azure App Service via automated GitLab CI/CD. Full data ingestion and retrieval pipelines with chunking strategy and query-time relevance filtering.

OCR-powered application automating warehouse label processing and matching workflows. Real-time label detection using computer vision with end-to-end backend pipeline for document automation and mobile frontend integration for operational use.