Skip to content

Open to AI engineering internships — remote, or Depok / Jakarta

I build AI agents that do real operational work.

I'm Vico — Information Systems at Universitas Indonesia, co-founder of Avagenc. I work on LLM agent orchestration, high-concurrency Go services, and serverless backends that keep running when nobody is watching them.

  • 3 systems in production
  • datafact.site live
  • Go · Python · TypeScript
  • AWS + GCP

Selected work

Three systems, and how they are put together.

Each one is a working product rather than a tutorial. The diagrams are drawn from the same data the text is, so they cannot drift apart.

Avagenc

2025 — present · Co-founder · engineering lead

In development

A multi-agent assistant that acts across your inbox, calendar, contacts, music and home devices from a single conversation.

In development. Go orchestration layer, Next.js web client and an Android client, containerised and deployed on cloud-native infrastructure.

  • Go
  • Next.js
  • Android
  • LLM agents
  • Vector database
  • Docker
  • GCP
One conversation, many tools. The orchestrator decides which to call and in what order.ChatGo orchestratorTool layerConnected servicesGmail · Calendar · Tuya

Datafact

2025 — 2026 · AI & backend engineer

Live

A serverless engine that turns a Google Form and a chosen persona into coherent, human-plausible survey responses at scale.

Live at datafact.site. Fully event-driven on AWS — nothing to keep warm between runs.

  • AWS Lambda
  • AWS Step Functions
  • API Gateway (HTTP)
  • EventBridge Scheduler
  • Generative AI
  • Serverless
Spiky load, no servers. Step Functions owns retries and partial failure; Lambda fans out one respondent at a time.Form + personaStep FunctionsLambda fan-outGenerative AIResponses

NusaVerify

2026 · Engineer — Bank Indonesia hackathon

Live

A fact-checker that scores how likely a claim is to be a hoax, and shows the entire reasoning chain behind the number.

Built for the Bank Indonesia hackathon. Live at nusaverify-web.vercel.app.

  • Next.js
  • TypeScript
  • LLM
  • Information retrieval
  • Vercel
Each source is scored on its own, then combined — so disagreement stays visible instead of averaging away.ClaimSource retrievalWeighted scoringVerdict + mind-map
Everything else I have built

How I build

Five layers, and what proves each one.

Listed by where it sits in a system rather than by language, because that is how the decisions actually group.

Agents & LLM

Getting a model to take actions reliably, and knowing when it has not.

  • Agent orchestration
  • Tool / function calling
  • Retrieval-augmented generation
  • Vector databases
  • LLM evaluation
  • Prompt design
  • n8n
  • Hugging Face

Proof Avagenc · Datafact · LLM evaluation — 200 questions graded for correctness, reasoning and hallucination

Backend & concurrency

Services that hold up when several things happen at once.

  • Go
  • High-concurrency services
  • REST API design
  • Third-party system integration
  • Distributed systems
  • Python

Proof Avagenc · gmail-sender

Cloud & delivery

Infrastructure that scales to a burst and costs nothing while idle.

  • AWS Lambda
  • AWS Step Functions
  • API Gateway
  • EventBridge Scheduler
  • Google Cloud Platform
  • Docker
  • CI/CD
  • Vercel

Proof Datafact · Avagenc

Interface

The part a person actually touches, including the explanation of what the system did.

  • Next.js
  • React
  • TypeScript
  • Tailwind CSS
  • Data visualisation

Proof NusaVerify · Vicoworks · HandlerIndonesia

Data & modelling

Measuring things, and training something when a rule would not do.

  • PostgreSQL
  • Supabase
  • Reinforcement learning
  • Gymnasium
  • Stable-Baselines3
  • Data analysis

Proof Optimizing Robot Tutor Strategies

Languages · Go · Python · TypeScript · JavaScript · Java · Dart · C

Research

Optimizing robot tutor strategies.

When should a robot tutor teach, rest, or push harder? Formalised as a 1,440-state Markov Decision Process and solved with deep reinforcement learning.

A reusable environment first, then four algorithms benchmarked against it across independent seeds.RobotTutor-v2RL agentLearned policy100% proficiency
  • A 1,440-state MDP over a 24-hour clock, learner proficiency, fatigue and engagement.
  • RobotTutor-v2 — a reusable Gymnasium environment, compatible with Stable-Baselines3.
  • DQN, PPO, TRPO and SAC benchmarked across independent seeds.
  • Soft Actor-Critic produced the best policy, reaching a 100% expert-proficiency rate and clearly beating random and fixed-schedule baselines.
  • Python
  • Gymnasium
  • Stable-Baselines3
  • DQN · PPO · TRPO · SAC

About

How I got here.

I started in Mechanical Engineering at Universitas Indonesia and moved to Information Systems after a year, once it was obvious that the systems I actually wanted to build were made of software. I am in my seventh semester.

Since February 2025 I have been co-founding Avagenc, an AI automation startup. In practice that means writing the agent orchestration layer, the Go services underneath it, and the deployment that keeps both alive — usually in the same week.

Before that I spent two months evaluating a mathematical LLM across 200 questions, grading correctness, reasoning validity and hallucination one answer at a time. It is the least glamorous work I have done and probably the most useful: it is where I learned what these models actually fail at, rather than what they are advertised to do.

I am looking for an AI engineering internship — remote, or on the ground in Jakarta and Depok.

Based in
Depok, West Java, Indonesia
Studying
B.Sc. Information Systems, Universitas Indonesia (2023 — 2027)
Looking for
Open to AI engineering internships — remote, or Depok / Jakarta

Contact

Let’s build something.

Open to AI engineering internships — remote, or Depok / Jakarta. The fastest way to reach me is email.