Hi, Iโ€™m Udochukwu, an AI engineer based in Lagos. I design and build AI applications, taking projects from early experiments through to deployment. My work spans system architecture, Python backends, model integration, evaluation, and the interfaces people use. I think carefully about where AI is useful, how to measure whether itโ€™s working, and what should happen when it gets something wrong. I care about shipping software that people can depend on.

View projects Rรฉsumรฉ

Projects

Lenslayer contract overview with a decision queue

Lenslayer

Built a contract intelligence and operations platform that carries agreements from review through negotiation and post-signature follow-up. It identifies risks and obligations, compares revisions, and answers questions with citations to the source clauses. Teams turn findings into approval decisions, assigned tasks, renewal and notice deadlines, and counsel handoffs.

Python ยท FastAPI ยท Next.js ยท TypeScript ยท Cloudflare

WHT recovery overview with a case queue and evidence coverage

WHT Recovery Control

Built an agent-assisted recovery workflow for Nigerian withholding-tax receivables. It ingests ledgers and receipts, uses DeepSeekโ€™s structured tool calls to extract fields with confidence scores and source quotes, then reconciles that evidence against invoice references, TINs, amounts and reporting periods. Exceptions move into a review queue for recognition, dispute resolution and closure. Cloudflare D1 and R2 retain case state, original evidence and audit history; consequential decisions stay with the reviewer.

DeepSeek ยท TypeScript ยท Cloudflare D1 + R2

ContractGuard

Built a document-intelligence workspace that ingests PDF, DOCX, TXT, and scanned documents, retrieves evidence with ChromaDB and all-MiniLM-L6-v2 embeddings, and uses Groq through LangChain to generate citation-linked risks, obligations, negotiation questions, and grounded follow-up answers.

Streamlit ยท LangChain ยท ChromaDB

Hybrid Retrieval

Built a retrieval service with FastAPI and PostgreSQL. It ingests and versions TXT, Markdown, and PDF sources, then combines ParadeDB BM25 with pgvector semantic search using Reciprocal Rank Fusion and page-level provenance. On SciFact (5,183 documents, 10,980 chunks), hybrid search reached 84.44% Recall@10 and 70.75% nDCG@10, against 79.07% and 67.73% for BM25 alone and 81.97% and 65.89% for vector search alone.

Python ยท FastAPI ยท PostgreSQL ยท pgvector ยท ParadeDB

AI engineering stack

Engineering deep dives

Explorations

Contact

Iโ€™m open to backend, AI application, and product-engineering opportunities where I can build reliable Python systems and ship customer-facing software.

Project preview

Screenshot from the public demo. Records shown are synthetic.