Professional Experience

HR / ATS ERP Platform

Role: Tech Lead

Dates: 2025

An agentic HR platform managing 30 employees, with document-grounded AI workflows for prospective roles and candidate analysis.

Internal project. Proprietary screenshots and implementation details are intentionally not published.

Context

During summer 2025, I led the architecture and implementation of an internal HR platform for Jointech AI with a team of five engineers. It manages 30 employees and supports analysis of prospective roles and candidates.

The problem

HR and candidate knowledge arrives as heterogeneous documents rather than clean database records. The platform needed to ingest that material, retrieve relevant evidence and give its AI workflows grounded context for evaluating roles and candidates.

What I did

• Led five engineers across technical architecture and implementation.
• Built the Node.js, Express and TypeScript REST API and its underlying data model.
• Containerized local and deployed environments with Docker.
• Implemented document ingestion, chunking, embeddings and Chroma vector search for grounded semantic retrieval across HR and candidate documents.
• Built agent orchestration and prompt-engineering workflows over MCP, connecting retrieval and structured tools to LLM calls.
• Developed model fine-tuning workflows and an LLM-as-a-judge evaluation step with prompt grading before release.
• Built the React client on top of those services.

Technical highlights

The retrieval pipeline turns heterogeneous documents into chunks and embeddings stored in Chroma, then supplies relevant evidence to agent workflows through structured tools. Evaluation and prompt grading provide a release check for LLM behavior rather than treating prompt changes as unreviewed configuration.

Outcome

Used internally in production to manage 30 employees and support hiring analysis.

Confidentiality

This is an internal product. Screenshots, proprietary implementation details, internal architecture and confidential data are intentionally not published.