---
title: Sulthan Zahran Ma'ruf, AI engineer
description: AI engineer in Indonesia building AI agents and the infrastructure they need in production: tool gateways, observability, and tests for agent-written code.
canonical_url: https://zahranm.cloud/
last_updated: 2026-10-10
---

# Sulthan Zahran Ma'ruf, AI engineer: I build what AI agents run on

Tool gateways, agent runtimes, and the traces and evals that keep them honest.

Now: AI agents at an AI software company · LG Sinarmas (smart factory)

Indonesia (UTC+7) · Rust · TypeScript · Go · Python · [CV (PDF)](https://zahranm.cloud/cv.pdf) · [GitHub](https://github.com/SulthanZahran1)

Skills: AI agents · tool use / MCP · RAG · LLM observability · agent harnesses · voice pipelines · mutation testing

## One agent turn, traced

The home page opens on one Bella turn drawn as a trace (timings and token counts are sample values). A WhatsApp message, "Move my 3pm with Andi to tomorrow, same time, and let him know.", flows through these spans:

1. sambungapi · whatsapp.inbound
2. bella · plan (model)
3. bella · calendar.findEvent (minRole allowlisted)
4. bella · role gate on calendar.update (minRole member)
5. google · calendar.update
6. bella · reply (model)
7. sambungapi · whatsapp.send
8. console · review → dataset

A toggle switches who is asking. As a member, the gate passes and the event moves. As an allowlisted guest, the gate refuses calendar.update, Google is never contacted, and Bella replies that moving events needs a member account. Scrolling (or clicking a span) follows the turn through three chapters. Each chapter is one of the systems I built, and every demo is a framed replica titled with its project name.

### 1. sambungapi: OAuth tool gateway for AI agents (Rust, primary engineer)

sambungapi is the Rust gateway I'm primary engineer on at an AI software company: a durable outbox for webhooks, and a token vault for agents' OAuth tool calls. Replica of private code, sample data.

- Inbound webhooks (e.g. WhatsApp message.received) are committed to a durable outbox in the same transaction and delivered at-least-once by a worker that retries with the same Idempotency-Key.
- Tool calls don't go through the outbox: each one is idempotency-keyed, unseals the AES-256-GCM OAuth token for that call only, and writes an execution_logs row. Tokens are refreshed before expiry, or once on a 401.
- Demo: the chapter plays this turn's path: the WhatsApp webhook is committed to the outbox, the first delivery to Bella gets a 503 and is retried with the same Idempotency-Key, and Bella's reply goes back out through whatsapp.send. Bella's own calendar calls go to Google directly, so other agents' OAuth tool calls are shown dimmed; pick one of seven providers to send one through the vault (the agent node switches to "other agent"), or try a tool call whose token is about to expire.
- Stack: Rust, axum, PostgreSQL, OAuth 2.0, AES-256-GCM, transactional outbox.

### 2. Bella: AI chief of staff (TypeScript agent worker, core engineer)

Bella is the AI chief of staff I'm a core engineer on: WhatsApp, Lark, Teams and web, one agent worker behind them. Every tool declares a minimum role, and a gate checks it before the call runs. Real tool names and roles, sample requests.

- Every tool declares a minimum role (admin, member, allowlisted), and a gate checks it before the call runs. A denial goes back to the model as `{ "ok": false, "verb": …, "error": { "code": "forbidden" } }`, and the model tells the user.
- Demo: the same message as the hero turn, checked by the gate: calendar.findEvent (minRole allowlisted) and calendar.update (minRole member). The asker is shared with the hero toggle; as an allowlisted guest, calendar.update is forbidden and the denial shown is the tool result the model receives.
- A harder request, "Book the sales sync for 10:00 tomorrow, and remind the team every Monday.", adds calendar.freebusy, calendar.create and workflow.create (minRole admin), so even a member is one role short.
- Stack: TypeScript, Hono, Go, MongoDB, Redis, LLM tool use.

### 3. Bella console: observability and evals (built most of it)

Most of Bella's observability console is mine: the waterfall, per-span tokens and cost, and a review loop that turns verdicts into eval cases. Real span attributes and dataset names, sample values.

- Demo: inspect any span of the turn (token usage for model spans, minRole for tool calls, the gate decision, the outbox path for sambungapi spans), then review it: correct → confirmed-routing, wrong skill → hard-routing, bad reply → hard-quality.

## Built solo, in the open

Open source and personal builds: trusting the code agents write, retrieval that weighs evidence, and real-time voice. Each card opens a full demo underneath the row.

### rust_mutant (and dart-mutant, gopher_mutant): would your tests catch a bug?

On crates.io. Plant small bugs; good tests fail. The card shows the six mutants at a glance: four killed, one survived (LOR), one equivalent (AOI). The playground is real rust-mutant 1.0.1 output (`--format json`) for a `fare(age, base)` function with two tests (adult, child).

- Before the missing test: MSI 80%. LOR `||` → `&&` survives, and the IR check (TCE) confirms it is real: the IR differs. After adding a toddler test: MSI 100%.
- AOI `/` → `/ 1 /` survives the tests but is marked `equivalent`: at opt-level=2 it compiles to the same IR hash, so it is excluded from the score. TCE runs only on survivors.
- 4 mutants (COR, LCR, RVR, AOD) don't compile and are excluded.
- Links: [rust_mutant](https://github.com/SulthanZahran1/rust_mutant), [crates.io](https://crates.io/crates/rust-mutant), [dart-mutant (Homebrew)](https://github.com/SulthanZahran1/dart-mutant), [gopher_mutant](https://github.com/SulthanZahran1/gopher_mutant), [how the IR check works](https://github.com/SulthanZahran1/rust_mutant/blob/main/crates/rust-mutant-tce/src/lib.rs).

### cv-search: search that weighs real skills

Retrieval, on request.

- Demo: the query "rust backend oauth integrations" over illustrative candidates, stepped through lexical + skill-field scoring (+8 per query term found anywhere, +18 when a term is a skill Gemini extracted into the structured profile), then an LLM rerank that rescores the lexical hits 0–100 with a reason.
- Stack: Go, React, Gemini extraction, LLM rerank.
- Links: [the scoring function](https://github.com/SulthanZahran1/cv-search-prototype/blob/main/api/main.go#L1228), [source](https://github.com/SulthanZahran1/cv-search-prototype).

### honjang: English → Korean, at the right speech level

Voice, live.

- Demo: pick a listener (a shopkeeper or your manager) and the same English sentence is rendered in 해요체 (polite, everyday) or 합쇼체 (formal); an auto mode picks between the two.
- Pipeline: it waits for Deepgram's UtteranceEnd (1.2 s of silence), translates the whole reply, then speaks it clause by clause. Next (ADR-0004): stream LLM tokens into TTS.
- Stack: Expo, FastAPI, WebSocket, Deepgram Nova-3, OpenRouter, ElevenLabs.
- Links: [live app](https://honjang.zahranm.cloud), [source](https://github.com/SulthanZahran1/honjang), [ADR-0004](https://github.com/SulthanZahran1/honjang/blob/main/docs/adr/0004-hybrid-llm-tts-streaming.md).

## Smart-factory software (LG Sinarmas)

LG Sinarmas, 2024 to now, EV-battery plant. MES, real-time dispatching and AMHS integration for a plant that runs around the clock, and a log analyzer that replays equipment handshakes as timing diagrams.

- Demo (illustrative incident): an AGV is stuck at EQ12. The analyzer replays the SEMI E84 handshake log (L_REQ and READY from the equipment, BUSY and COMPT from the AGV) as a timing diagram and diagnoses a TA3 timeout: READY took 94 s to drop after COMPT, so AGV07 was held at EQ12.
- Stack: MES, RTD, AMHS, C# / .NET, SQL Server, Python, KO–EN RAG.

## Career, as a timing diagram

Each lane is a track, each box a role; clicking one shows a one-line readout and the details. Lanes, top to bottom: AGENTS (an AI software company), BUILDS (personal AI builds), FACTORY (LG Sinarmas, including the Korea deployment), LAB (FAAN Lab), FIELD (PLC internship, instrumentation engineer), EDU (B.Sc. Physics). Full history in the [CV (PDF)](https://zahranm.cloud/cv.pdf).

- **Software Engineer, AI agents.** An AI software company, part-time, remote.
  - Built sambungapi in Rust (axum): an OAuth tool gateway with an AES-256-GCM token vault, multi-org tenancy and a durable outbox, through which agents act on Google, Microsoft, GitHub, Slack, Notion, Linear and Jira.
  - Core engineer on Bella's agent worker (TypeScript): the runtime that plans, calls tools (Gmail, Calendar, Lark and more) and replies on every turn, over WhatsApp, Lark, Teams and web.
  - Built the messaging engine between apps and chat platforms, and the observability behind it: per-turn cost, execution waterfalls, and a human-review loop that produces ground-truth datasets.
- **AI builds, shipped.** Personal, open source, self-hosted. Jun 2026 to now.
  - lgensol-wiki (Jun): an LLM-compiled wiki with chat and a knowledge graph over public LGES posts.
  - cv-search (Jun): résumé search with Gemini extraction, deterministic scoring and an LLM rerank.
  - honjang (Jul): real-time English → Korean voice translation that picks the right honorific level.
  - rust_mutant, dart-mutant, gopher_mutant (Aug): mutation testing for the tests coding agents write.
  - recruit (Aug): the pitch copilot below, grounded only in my work profile.
- **Software Engineer, smart factory.** LG Sinarmas Technology Solutions, Karawang. Apr 2024 to now.
  - MES, real-time dispatching (RTD) and AMHS integration for an EV-battery plant: C# / .NET, WPF, SQL Server and Oracle.
  - Wrote a Python + SQL log analyzer that replays AMHS events as timing diagrams for incident diagnosis.
  - Built a Korean–English RAG pipeline (LangChain, PostgreSQL) over factory documentation, and automated monitoring across the plant's IoT sensor web interfaces.
- **AMHS Integration Specialist.** LG Energy Solution Ochang, South Korea. Aug to Oct 2025.
  - Integration-tested AGV, conveyor and stocker subsystems against the MES before production, and fixed the integration bugs in code.
- **Instrumentation Engineer.** PT Polychemie Asia Pacific Permai. Jan to Mar 2024.
  - Prototyped an ADC-based monitoring rig so temperature and flow could be logged on a PC instead of by hand; evaluated DAQ, PLC and Arduino options and wrote the plotting GUI.
- **Research Assistant, software.** FAAN Laboratory, Physics, Universitas Indonesia. Oct 2023 to Mar 2024.
  - Built the lab's MATLAB spectroscopy app: real-time acquisition, peak detection and fitting, Savitzky–Golay filtering and Tauc plots.
- **PLC Intern.** PT Sugitama Intiarto. Jun to Jul 2022.
  - First contact with industrial control software.
- **B.Sc. Physics, Instrumentation.** Universitas Indonesia. Aug 2019 to Jul 2023.
  - IndySCC22 student cluster competition (HPC): 4th place, the only team from Southeast Asia.
  - Programming lead for the UI Robotics Team at KRTMI.

## Your visit, traced

Every demo a visitor touches emits a span, and the page draws the visit as a waterfall, like a Bella turn, coloured by span kind (page, gateway, model, gate, tool, human, failure), with a playful "review this visit" step. It is a collapsed bonus at the end, clearly separate from the work. The trace is kept in the tab's memory only; nothing is stored or sent anywhere.

## Contact

Building agents that touch real systems?

- Email: zsulthan9@gmail.com
- CV (PDF): https://zahranm.cloud/cv.pdf
- GitHub: https://github.com/SulthanZahran1
- LinkedIn: https://www.linkedin.com/in/sulthan-zahran-ui

Recruiter? [Paste a job description](https://zahranm.cloud/recruit/) and get a short, evidence-based pitch from my work profile ([source](https://github.com/SulthanZahran1/jd-pitcher)).

## Index: everything running on zahranm.cloud

Self-hosted on one VPS behind Traefik (a collapsed list at the bottom of the home page). Gated demos are available on request.

| Property | Description | Status | Source |
| --- | --- | --- | --- |
| [honjang](https://honjang.zahranm.cloud) | English–Korean real-time voice translator with honorifics | live | [source](https://github.com/SulthanZahran1/honjang) |
| [recruit](https://zahranm.cloud/recruit) | LLM pitch generator: job description in, evidence-grounded pitch out | live | [source](https://github.com/SulthanZahran1/jd-pitcher) |
| [relay](https://relay.zahranm.cloud) | AI-native CRM on an event-driven workflow engine | live | |
| [teach](https://teach.zahranm.cloud) | Course: build an agent harness from scratch | live | |
| [wiki](https://wiki.zahranm.cloud) | LLM-compiled wiki with chat and a knowledge graph, from 556 public LGES posts | on request | [source](https://github.com/SulthanZahran1/lgensol-wiki) |
| [cv-search](https://cv-search.zahranm.cloud) | Résumé search: Gemini extraction, lexical + skill scoring, LLM rerank | on request | [source](https://github.com/SulthanZahran1/cv-search-prototype) |

Other builds:

| Property | Description | Status | Source |
| --- | --- | --- | --- |
| [emrisearch](https://emrisearch.zahranm.cloud) | GUI for extreme-mass-ratio-inspiral search runs | live | [source](https://github.com/SulthanZahran1/emrisearch-gui) |
| [aquaflow](https://aquaflow.zahranm.cloud) | Prepaid smart water metering, three roles | live | |
| [dcim](https://dcim.zahranm.cloud) | Data-center infrastructure management on NetBox + React | on request | |
| [hkbp](https://hkbp.zahranm.cloud) | Church admin system: Laravel → Vue + Go, self-hosted OIDC | on request | [source](https://github.com/SulthanZahran1/hkbp-jatinegara) |

## Sitemap

- [Home](https://zahranm.cloud/)
- [About](https://zahranm.cloud/about)
- [Contact](https://zahranm.cloud/contact)
- [Privacy](https://zahranm.cloud/privacy)
- [Recruiter copilot](https://zahranm.cloud/recruit/)
- [llms.txt](https://zahranm.cloud/llms.txt)
- [XML sitemap](https://zahranm.cloud/sitemap.xml)