Skip to content
malamsyah

Mochammad AlamsyahAvailable for new engagements · ~20 hrs/week

I fix LLM infrastructure costs and productionize agent systems —fractionally.

Software engineer in Tokyo, nine years building consumer-scale platforms: a super-app at scale, a data-warehouse bill cut 90%, production AI on the Claude Agent SDK and AWS. Platform judgment for LLM-heavy teams, without the headcount.

Start a conversation See the work

Blueprint of an LLM request path: a client hits a router, which either returns from cache or sends the miss to a model, then through an eval harness, before a response. Cost is measured on the model path.

CACHE
hit path first
ROUTE
batch vs realtime
EVAL
measure before ship
COST
bill tracks usage
Lat 35.6762° N — Lon 139.6503° E · GMT+9SysNominal

[ 01 — What I take off your plate ]

  • 01

    LLM spend, measured down

    Caching, routing, model right-sizing, batch vs realtime. Your bill should track usage, not outgrow it — the same discipline that cut a company-wide data-warehouse spend by 90%.

  • 02

    From demo to boring reliability

    Retries, guardrails, state, observability — agent pipelines and retrieval built with the discipline of any other distributed system. The gap between a great demo and a system nobody worries about is exactly the work I do.

  • 03

    An eval harness, not a gamble

    Every prompt or model change is a coin flip until there's a harness. I build the measurement that turns changes into measured rollouts.

  • 04

    Distributed systems & SRE

    Go, Kubernetes, AWS and GCP. Monolith decomposition, incident response, alert hygiene — the unglamorous plumbing that keeps products alive.

[ 02 — The work ]

Nine-plus years across fintech, ride-hailing, healthcare, and AI — leading distributed teams across Japan, India, Indonesia, and US time zones. Names withheld; specifics available in conversation.

  1. 012017 — 2018

    withheldcompany name withheldE-commerce startup — Jakarta

    Co-founder & CTO

    • Built the MVP of a bidding-based e-commerce platform across web and iOS
    • WebSocket bidding engine and multi-gateway payment integrations
    • Go, React, Node.js, Swift, Redis, Amazon SQS

    FounderGoWebSocket

  2. 022018 — 2023

    withheldcompany name withheldRide-hailing super-app — Bangalore & Jakarta

    Lead Software Engineer

    • Decomposed a core monolith into microservices; helped migrate primary services from VMs to Kubernetes
    • Built a Go distributed-locking library on Redis Cluster; cut alert volume from 30+ a week to under 5
    • Designed the ride-hailing homepage on a BFF architecture; shipped a phone-number masking feature later adopted org-wide

    GoKubernetesDistributed systems

  3. 032023 — 2025

    withheldcompany name withheldFinancial-services conglomerate — Jakarta

    Lead → Principal, AI & Data Platform

    • Built an AI assistant for C-level executives: LLM function-calling over the company data warehouse for natural-language business metrics
    • Migrated the company-wide data warehouse to cloud-native serverless analytics — 90% cost reduction
    • Architected and led a loan-application system for the used-vehicle market on GCP, development through production

    AI platformBigQueryGCP

  4. 042023 — present · concurrent

    withheldcompany name withheldGlobal healthcare platform — remote

    Senior Software Engineer (Consultant)

    • Production AI tooling on the Claude Agent SDK, AWS Lambda, and Go
    • Key contributor scaling the platform from a single market to dozens of countries — multi-currency, multi-language, multi-payment-gateway
    • Prototyped a bulk-upload validation engine that turned a ticket-analysis task into a working MVP

    GoClaude Agent SDKAWS

[ 03 — Proof ]

90%Read the case study

Group-wide data warehouse, re-architected

Apache Hive on ~100 VMs to an S3 data lake + Redshift Serverless — ≈$20–30k/month down to ≈$2k, half the scheduled reports retired first. Three months, zero cutover incidents.

Monthly bill
$20–30k → $2k
Timeline
3 months
Scheduled reports
30 → 15
VMs retired
~100
100M+Read the case study

Super-app home screen on one BFF call

Six-plus API calls collapsed into one Go service — 100M+ calls/day at <50ms p90, with parallel fan-out, per-service timeouts, and a zero-incident migration.

Calls per screen open
6 → 1
Daily volume
100M+
Latency
<50ms p90
Migration
~1 mo to 90%

Built in the open

Open source and certifications

Google Cloud Professional ML EngineerCertified Kubernetes AdministratorAWS Solutions ArchitectGoogle Cloud Associate Engineer

[ 04 — How I work ]

Ways of working

Advisory

A standing line to a senior platform engineer. Architecture reviews, cost audits, hiring help, written decisions you can forward to the team.

Embedded fractional

Hands in the codebase a few days a week. I ship the platform work alongside your team and leave the runbooks behind.

Fixed scope

A defined deliverable — an eval harness, a cost-reduction pass, a productionized agent pipeline — with a written scope before any contract.

Paid pilot · Every engagement starts with a short, paid pilot: tight scope, written findings, a decision at the end. No pressure to continue — the pilot has to earn it.

  1. 01

    Async-first, from Tokyo

    A track record of leading distributed teams across Japan, India, Indonesia, and US time zones. Async hours plus one weekly call — my mornings overlap US afternoons, so you get written decisions, not meetings.

  2. 02

    Evals before scale

    Nothing ships to more traffic until there's a measurement that says it should.

  3. 03

    Boring is the goal

    Production AI systems should be the least exciting part of your stack.

  4. 04

    Leave the runbooks

    Every engagement ends with your team able to run what I built without me.

Currently taking conversations

Tell me what's breaking — or what's about to. I reply within a day with three questions and an honest read on fit; if it makes sense, we scope a short paid pilot on a call.

hi@malamsyah.com

References available on request · github.com/malamsyah