Build and own the core of Agent using the AI Agents SDK (TypeScript) and the hosted Agents API- agent instructions, tool orchestration, and per-conversation sessions.

Design the agent’s decision-making so it reliably understands requests in Chat and selects the correct tools to act on them.

Build document-understanding pipelines that extract requirements from files in any format, including scanned documents and images via OpenAI vision.

Implement retrieval (RAG) using hybrid search - pgvector embeddings plus PostgreSQL full-text - with sound chunking, versioning, and relevance tuning.

Define and maintain the AI evaluation framework: build eval sets from past projects, set per-wave accuracy targets, and enforce the evaluation gate in CI so quality never regresses.

Implement guardrails and prompt-injection defenses, ensuring all documents and messages are treated as data, never as instructions, and that the model performs language tasks only - never computing numbers, dates, or permissions.

Deliver multilingual understanding, including mixed-language conversations, with consistent, correct responses.

Continuously review agent accuracy against saved guidance, and improve the system from real usage.