Heliodoron is an AI studio building vertical AI products for Indonesian industry. This page is how that actually works — the discipline, the engineering stack underneath, and the economics that make it viable across many industries.
We research many. We pick one. We embed with the people who do the work, learn the workflow that actually happens (not the one on the org chart), and build the AI product that fixes the pain — then we operate it so the client never has to.
The discipline is in the sequencing: research across the Indonesian economy, ship one vertical at a time, go deep enough to build something the industry actually needs, and stay long enough to operate it as a recurring service. Heliodoron isn't a horizontal platform. It isn't a consultancy. It's a studio that researches widely and ships narrowly — one vertical at a time, with the next already on the bench.
The studio runs on a shared engineering stack. Each vertical product we ship inherits it. Each new product hardens it for the next.
Multi-step AI workflows that chain models, handle retries, and manage execution state.
Whatever the vertical produces — drawings, documents, photos, voice — parsed into structured records the AI can act on.
The glue that fits a general-purpose foundation model to a vertical's vocabulary and outputs.
How we measure whether a model output is correct for a given vertical's standards.
A white-label chassis where the widgets change per vertical, but the framework is one codebase.
Every deployment improves the next one. Bugs found in vertical one harden the stack for vertical two.
No customer ever buys these directly. They're internal capital equipment. But every vertical product Heliodoron ships runs on them — which is why the second vertical launches in weeks, not years.
The most expensive part of getting to orbit was discarded after a single flight. The economics inverted overnight after the arrival of reusable rockets.
AI delivery has the same waste. Most consultancies rebuild the orchestration, the evaluation framework, and the deployment pipeline from scratch on every project. The means of production gets thrown away with each engagement. The math never works.
New vertical every launch. Same stack. Same engineering team.
Heliodoron doesn't throw the stack away. Each new vertical inherits everything the last one built. This is what makes a small, disciplined team economically viable across many industries — not shortcuts, not generic tools repackaged with different logos, but an engineering stack that compounds every time it's deployed.
The compounding isn't metaphorical. Four multiplicative factors decide whether any vertical we pick becomes a business. The framework is identical across industries — only the numerator and denominator move.
Frontier-model APIs price input at $0.075–0.15 per million tokens. A typical user interaction costs $0.0005–0.003. An active user generating 100 interactions a month costs $0.05–0.30 in API spend. Intelligence is now a commodity input.
Local pricing for SMB and professional tools sustains IDR 49,000–499,000 per seat per month ($3–30 USD). After API, infrastructure, and support, gross margin lands above 80%. Consumer-grade economics on B2B tools.
65 million UMKM. 213 million internet users. Vertical subsegments count in the hundreds of thousands each. Sub-percent penetration on a single vertical is a real business. Multiple verticals on one stack is a portfolio.
First vertical: months of build. Second: weeks. Third: days of vertical-specific work layered onto a hardened stack. Engineering investment paid once, recovered N times. The reusable rocket, applied to AI delivery.
Multiply those four — cheap inference, sustainable local pricing, deep vertical markets, compounding stack — and the math finally works. Not for one-shot consulting engagements, not for English-only SaaS importing US pricing into Indonesia, but for vertical-specific AI products built and operated locally, at scale.
Across Indonesian sub-verticals at scale — from private business to 65M UMKM to regulated enterprise — we've scoped where vertical AI fundamentally changes the workflow. The first vertical was selected by depth of pain we observed in the field, not market-sizing decks. The second is already on the bench. Each product launches publicly only when it's ready to operate at scale; until then, the work happens quietly and the stack compounds.
One concrete external signal so far: our prototype NetPulse AI — an agentic system that turns Indonesian telecom customer complaints into structured incident tickets in under 15 seconds — has been selected into the Top 100 of Google Cloud Gen AI Academy APAC 2026 · Cohort 1. Built on Google ADK, BigQuery, AlloyDB AI, and Vertex Gemini.