Pangaea: Taking AI from Pilot to Production

A production AI practice that moves systems past the pilot stage — retrieval-augmented generation gated on measurable quality, red-team-tested for safety, and built to run reliably in client environments.

RAPTEK TeamPangaea Digital Labs
  • Artificial Intelligence
  • Data Analytics
  • Software Development
  • AI Safety

Delivered by Harry Osmar Sitohang

View live project
Pangaea: Taking AI from Pilot to Production

Overview

Most AI projects stall in the same place: a demo that dazzles in a meeting but cannot be trusted in production. Pangaea was built to close that gap — an AI implementation practice that takes systems from pilot to production, where they answer real users reliably, safely, and at scale.

This is the latest work of our Senior Principal Engineer, Harry Osmar Sitohang, who designed and built it end to end. It is the discipline he now brings to RAPTEK’s own AI and data engagements.

What was delivered

  • Production RAG systems. Retrieval-augmented generation that grounds answers in a client’s own knowledge — not the model’s guesswork — so responses are accurate, current, and traceable to a source.
  • Evaluation gated on measurable quality. Every answer is scored before it ships. Faithfulness is held to a ≥ 0.90 threshold (using RAGAS-style evaluation), so a release has to prove it is grounded rather than merely look convincing.
  • Safety by red-teaming. Systems are adversarially tested against jailbreaks, prompt injection, and PII leakage before they reach users — measured against the OWASP Top 10 for LLM Applications and the NIST AI Risk Management Framework.
  • Secure-by-design architecture. Client data stays inside the client’s own environment, so adopting AI never means handing sensitive data away.
  • The software around the model. Production AI is mostly software: the agents, CI/CD pipelines, and observability tooling that keep a system shippable and debuggable long after launch.

How it works

Pangaea’s pipeline: retrieval grounds the model in the client’s own knowledge, an evaluation gate scores every answer for faithfulness and blocks anything below 0.90, red-team tests check for jailbreaks, injection and PII leakage, and only then does a response reach the user

The gate is the point: nothing reaches a user until it has proven it is grounded and safe.

The difference between a pilot and a production system is the gate in the middle. A pilot ships whatever the model says; a production system measures the answer first and refuses to ship the ones that fall short. By making quality and safety into numbers that block a release — not hopes that ride along with it — the same discipline that protects a banking API gets applied to a language model.

The result

Pangaea serves teams across Indonesia and Southeast Asia — from high-growth startups to enterprises including Golden Rama and TripDeals — and the same practitioner shipped the national-scale work behind Kartu Prakerja. The throughline is a refusal to confuse a convincing demo with a dependable system.

If you have an AI pilot that needs to become something you can trust in front of real users, that is exactly the work our Data Analytics and Software Development teams do. Start a conversation.

Share this page

Portfolio
Free consultation