Iman Suherman

suherman.net

Build real products with AI agents — from idea to running application.

About

A hands-on engineering leader

Iman Suherman is a technology leader with 20+ years designing and scaling cloud systems, microservices, and modern web/mobile platforms. He is currently Principal Engineer at Qantas Loyalty (Universal Retail Platform) and DevOps Engineering Lead at Mekari.

After hours is the playground: he spins up real products on suherman.net — macOS and mobile apps, developer tools, local AI gateways — then signs the builds, wires the registries, routes traffic through Cloudflare, and lets AI agents ship from a clear SPEC instead of vibes.

About →

Webinars

Learn from live Saturday builds

Free sessions on shipping with AI agents — recordings, agendas, slides, and notes anyone can study. Watch both Saturday builds below.

Featured: Train your LLM / agent (19 Sep, Bahasa) and Build an App with AI, Live (12 Sep, English). Start with a recording, then skim agenda and resources.

Recordings

Watch the sessions

Cara Melatih LLM / Agent Anda: Dari Prompt sampai Persona

Sabtu, 19 September 2026
11:00 – 12:00 WIB
12:00 – 13:00 Singapore / Malaysia
14:00 – 15:00 Sydney

Live Meet ended · recording published

Open slides (PDF)Read session notes

Agenda

What we covered

01 · 10 min
What “train” means for LLMs vs agents
Prompting, evals, RAG, fine-tuning, and persona behaviour — which lever fits which problem.
02 · 15 min
Prompts, evals & contracts
System prompts, acceptance checks, and treating behaviour as a reviewable contract before touching weights.
03 · 15 min
RAG vs fine-tune (and when not to)
Knowledge injection vs behaviour change — cost, data, and failure modes with a practical decision frame.
04 · 10 min
Local LLMs & persona agents
Ollama / MLX for cost and control; teaching an agent to emulate a user persona without leaking secrets.
05 · 10 min
Live demo + Q&A
TripSpec live: a Japan plan, three flights, no prices — then questions.

Outcomes

What we take forward

Harness-first training mindset, a published recording, and themes for the next deep-dive from the audience survey.

Train the harness, not the weights
Specs, system prompts, evals, and guardrails reshape behaviour without swapping the model file — the core message of this Saturday.
Recording + notes published
The Bahasa session is on YouTube with on-site session notes so anyone who missed the Meet can still follow.
Local models need memory headroom
Ollama demos across platforms work until GPU memory runs out — plan VRAM when choosing local models.
Weekly Saturday sessions (free & open)
Next Saturday opens when scheduled — vote topics in the survey so we know what to deep-dive.

Slides, session notes, and the TripSpec repo are linked on the right.

What we built from

Top themes from the pre-session poll

These votes steered the training deep-dive.

01Prompt engineering & system prompt6 votes
02LLM lokal (Ollama / MLX)2 votes
03Cost & latency: cloud vs lokal1 votes
04Eval / acceptance untuk agent1 votes
05Guardrail & safety untuk agent1 votes

Covered through TripSpec — system prompts first, then local models, without changing model weights.

Resources

Demo slides

TripSpec recipient PDF from the live session.

TripSpec-2026-09-19.pdf

Repository

TripSpec source — open anytime.

https://github.com/iman-suherman/sdd-travel-planner

Earlier sessions

Previous Saturday live builds — watch, skim the agenda, and reuse the ideas.

PastEnglish≈ 94 minutes (recording)

Build an App with AI, Live: From Idea to Running Application

Completed · English · SurveyDesk live · recording on YouTube

Saturday, 12 September 2026
11:00 AM – 12:00 PM WIB
12:00 PM – 1:00 PM Singapore / Malaysia
2:00 PM – 3:00 PM Sydney

Recording: https://youtu.be/h3FPQEw2QNI

Session notes: Open session notes

We built SurveyDesk from Spec live — Spec-Driven Development, Agent on Rails waves, and a local-first survey desk on phone, tablet, and web. Watch the recording.

01 · ~15 min
Spec-Driven Development as the contract
AI as an engine under human control — Gherkin / BDD specs and Architecture Decision Records (ADRs) as guardrails against hallucination.
02 · ~25 min
SurveyDesk live demo
Local-first SQLite survey desk — native iOS/Android + web, JSON form specs, draft → open → closed lifecycle; validated on phone, tablet, and simulators.
03 · ~15 min
Agent on Rails orchestration
Parallel agents in execution waves, Gherkin acceptance checks, and specs as the control plane so agents update features without stepping on each other.
04 · ~5 min
Local agents & what’s next
Why run coding agents locally for cost and control. Next Saturdays: Bahasa session, then training agents to emulate user personas.

SPECSurveyDeskAgent on RailsEnglishrecordingcompleted

PastBahasa Indonesia60 minutes

Build an App with AI, Live: From Idea to Running Application

First live session (Bahasa Indonesia) — completed

Saturday, 5 September 2026 · 11:00 – 12:00 WIB (Western Indonesian Time)

First live session completed — thank you for joining. Feedback shaped clearer language, recordings, and the English follow-up on 12 Sep.

01 · 10 min
Context & concepts
Problem → research → why we do not immediately tell the AI “build the app”.
02 · 5 min
Turn the idea into a SPEC
Spec as contract: routes, behaviour, copy, components, hard rules, acceptance criteria. The prompt is not the source of truth.
03 · 25–35 min
Kick off AI agents live + narrate
Orchestrator splits work (foundation → parallel pages → i18n/polish → typecheck). While agents run: ownership, knowledge, preventing hallucination, when humans decide.
04 · 10 min
Resume demo & evidence
Local URL, typecheck/test, smoke walk — runnable evidence, not just a prototype from a prompt.
05 · 15 min
Product factory + Q&A
Why many apps: each one is a product experiment plus brand exposure. What scales is the development system.

SPECAI agentslive buildBahasa Indonesiacompleted