[01] | PORTFOLIO — 2026
AGENT_EVALS · PASS
Engineering AI systems that act. reason. assist.
Engineering
AI systems
that
act.
reason.
assist.
I’m Prince, an AI Engineer building intelligent systems and automation to solve real human problems. I’m particularly interested in building AI agents that can reason, take action, and be trusted in production.
AGENT_EVALS · PASS
[02] | ABOUT
Prince N. Oboe-Sam
AI Engineer · Agentic Systems
I started my career in Digital Marketing, but my curiosity gradually pulled me deeper into AI. What started as an interest in marketing AI automations has grown into a serious focus on building AI systems.
Today, I’m focused on building agentic AI systems that connect LLMs, data, and business workflows to solve real problems. I enjoy exploring how AI can move from simple automation to systems that can reason, interact with external services, and perform useful tasks.
I’m particularly interested in building reliable AI agents that go beyond generating responses to understanding context, taking action, and being evaluated for how well they perform in real-world situations.
- Experience
- AI Customer Support Agent
Built a customer support agent for EliteCare Medical Center that handles patient inquiries, consultation payments, and cal.com appointment booking.
- Ad Metrics Automation
Built an automated Facebook ad metrics reporting workflow that sends performance metrics to Google Sheets, customer support agents for dental clinics on the Lukins Donkomi marketing team.
[03] | PROJECTS
Featured projects.
01
● AI AGENT
Lead Qualification AI Agent
An AI-powered WhatsApp agent that qualifies leads, captures them into a google sheet tool, and answers business-related questions.
- - Agent orchestration: Built the agent with LangChain to orchestrate LLM reasoning, retrieval, conversation state, and tool execution across the customer workflow.
- - Human-in-the-loop: Implemented human takeover with both manual and automatic escalation, allowing staff to take over conversations directly while enabling the agent to disable itself and hand off to a human when human intervention is required.
- - Hybrid RAG retrieval: Combined semantic vector search (0.5) with lexical BM25 search (0.5) and used Reciprocal Rank Fusion (RRF) to improve retrieval across both meaning-based and exact keyword matches.
- - LLM evaluation: Built single-turn and multi-turn evaluation datasets to measure agent correctness and behavioral alignment across different conversation scenarios.
- - CI/CD pipeline: Automated agent evaluations on pull requests, enforced regression checks before merging, and deployed approved changes to production through Railway.
Pinecone
Docker
Github Actions
FastAPI
LangChain
LangSmith
OpenRouter
Railway
Python
02
● RAG SYSTEM
Legal Documents RAG
An end-to-end legal RAG system built around the 1992 Constitution of Ghana, combining constitution-aware document processing, hybrid retrieval, metadata filtering, cross-encoder reranking, grounded generation, and automated evaluation. Achieved 90% faithfulness and 82.7% answer correctness across a 37-question evaluation set.
- - Built an end-to-end RAG pipeline from document parsing and metadata extraction to retrieval, reranking, and grounded generation.
- - Engineered hybrid retrieval combining semantic search, BM25, metadata-aware retrieval, and weighted Reciprocal Rank Fusion.
- - Implemented query routing and cross-encoder reranking to dynamically select retrieval strategies and prioritize relevant legal context.
- - Designed grounded generation with few-shot prompting and explicit abstention controls to improve answer consistency and reduce hallucinations.
- - Built automated RAG evaluation and CI regression gates using Ragas and custom safety checks.
Evaluation Results
| Metric | Score |
|---|---|
|
Faithfulness |
90% |
|
Answer Correctness |
82.7% |
|
Context Precision |
98.3% |
|
Context Recall |
100% |
|
Safe No-Answer Rate |
100% |
|
Evaluation Set |
37 ques. |
Pinecone
Hugging Face
LangChain
OpenRouter
Ragas
Python
Github Actions
03
● AI INFERENCE
Document Intelligence
An end-to-end document intelligence system that extracts structured information from receipt images using OCR, adaptive preprocessing, and Vision LLM pipelines, evaluated across 361 SROIE receipts.
- - Multi-pipeline extraction: Built and compared OCR + LLM, adaptive OCR + LLM with CLAHE fallback, and direct Vision LLM architectures for extracting company, address, date, and total.
- - Evaluation: Evaluated 361 SROIE receipts, achieving 0.858 overall F1 with adaptive OCR + LLM and 0.854 overall F1 with Vision LLM, while measuring field-level accuracy and inference latency.
- - Production API: Built a FastAPI service with Pydantic schema validation and automated tests, exposing OCR and Vision LLM extraction through a unified REST endpoint.
- - Deployment: Containerized the system with Docker and deployed both extraction pipelines to Railway, making the document extraction service available as a production API. ChatGPT can make mistakes. Check important info.
FastAPI
Hugging Face
Pydantic
OpenRouter
Docker
Python
Railway
[04] | STACK
My tech stack.
Technologies I use to build, evaluate, and deploy AI systems that work in the real world.
- LLM & Automation
- - Claude Agent SDK
- - Hugging Face
- - OpenRouter
- - Pydantic AI
- - LangChain
- - n8n
- Backend & Dev
- - TypeScript
- - FastAPI
- - Python
- - SQL
- Eval & Observability
- - LangSmith
- - Ragas
- Vector Store
- - ChromaDB
- - Pinecone
- DevOps & CI/CD
- - Github Actions
- - Docker
- - Git
[04] | STACK
My Tech Stack.
Technologies I use to build, evaluate, and deploy AI systems that work in the real world.
- LLM & Automation
- - Hugging Face
- - OpenRouter
- - LangChain
- - n8n
- Backend & Dev
- - Pydantic
- - FastAPI
- - Python
- - SQL
- Eval & Observability
- - LangSmith
- - Ragas
- Vector Store
- - ChromaDB
- - Pinecone
- DevOps & CI/CD
- - Github Actions
- - Docker
- - Git
OPEN TO WORK
[05] | CONTACT
Let's build something
{intelligent}
Currently open to staff-level roles and serious
freelance. The fastest way to reach me is email.