[go: up one dir, main page]

Open to Solutions Architect & CTO roles

Jack LeCTO · Solutions & Technical Architect

I architect platforms that stay correct under load — and lead the teams that ship them.

18 years from multi-threaded telecom systems to AI-native products. I sit between the business and the engineers: run the solution workshop, draw the C4, write the ADRs, then build and load-test the first version myself.

Based in
Vietnam · on-site in USA, UK, Australia
Languages
English · Vietnamese
Reference architectureClients pass through an edge layer and an API gateway into domain services, an AI orchestration layer and an event ledger, each backed by its own data store, all inside a private Kubernetes network.GKE / EKSPRIVATE VPCClientsweb · mobile · AI agentsEdgeCDN · WAF · signed ticketsAPI gateway / BFFauthn · quotas · idempotencyDomain servicesNestJS · FastifyAI orchestrationLLM · OCR · agentsLedger & eventsappend-only · outboxPostgreSQLOLTP · row securityVector storeQdrant · pgvectorRedis · queuecache · BullMQ
fig. 01 — reference architecturerequest path · live
  • 18

    years shipping production systems

    since 2008

  • 300

    engineer division I was Technical Architect for

    TMA Solutions · DG1

  • 96.0%

    recall@5 on LongMemEval-S, 500 questions

    Cortex Hub benchmark

  • 10,500+

    players in the first month after launch

    Xóm Nông Trại · Oct 2026

  • 4,000

    CCU on 4 shards — one 4-vCPU box that also runs Postgres and Redis

    Xóm Nông Trại load test

  • 4

    countries on-site with enterprise clients

    Vietnam · USA · UK · Australia

How I take a system from a whiteboard to production.

The same four moves whether the client is an insurer, a bank, a game studio or a farm operator — and each one leaves an artifact behind that the team can hold me to.

  1. 01

    Frame the problem with the people who own it

    Solution workshops and pre-sales with insurers (AAA Insurance), banks and hospitals. I turn a business outcome into constraints a system can be judged against: volumes, latency budgets, data residency, audit and compliance.

    • Stakeholder discovery
    • Non-functional requirements
    • Build vs. buy
    • Cost envelope
    example · solution-brief.mdbrief
    goalcut claim turnaround from days to hours
    loadmany insurers · month-end bursts
    latencyp95 < 500 ms on submit
    dataPII stays in-region, audit every read
    aimodel drafts · human decides
    step 01 · artifact
  2. 02

    Model it so everyone can argue with it

    C4 context → container → component diagrams, ADRs for every irreversible decision, explicit data contracts. Diagrams that a COO and a junior engineer read the same way.

    • C4 diagrams
    • Architecture Decision Records
    • Data contracts
    • Threat model
    adr-0007-ledger.mdadr
    statusaccepted
    contextevery currency change must be explainable
    decisionappend-only ledger + idempotency keys
    alsotransactional outbox for side effects
    tradeoffmore rows, zero silent drift
    step 02 · artifact
  3. 03

    Build the platform, not just the feature

    Service boundaries, golden paths and shared rule libraries so the client predicts and the server decides. LLM, OCR and agent workflows wired in behind approval gates — the model proposes, a human signs off.

    • Microservices & APIs
    • LLM / OCR / agent integration
    • Human-in-the-loop
    • GCP · AWS · Kubernetes
    ops-agent.flowflow
    read26 tools · no DB, no SSH, no SQL
    write3 capped tools · daily quota
    propose11 tools → admin_approvals
    approvehuman, never the requester
    auditevery action, with a reason
    step 03 · artifact
  4. 04

    Prove it with numbers, then operate it

    Load tests with interleaved A/B runs, parity gates that fail the build on a single bit of drift, p99 budgets wired into dashboards. A change ships when the benchmark says so.

    • Load & soak testing
    • SLOs & observability
    • Profiling
    • Capacity planning
    ra2-bench · 64 playersbench
    tick p501.6 ms
    tick worst8.7 ms · none > 15 ms
    bandwidth0.8 KB/s per player
    dropped0
    paritynative ≡ wasm to 1e-9
    step 04 · artifact
example · solution-brief.mdbrief
goalcut claim turnaround from days to hours
loadmany insurers · month-end bursts
latencyp95 < 500 ms on submit
dataPII stays in-region, audit every read
aimodel drafts · human decides
step 01 · artifact
adr-0007-ledger.mdadr
statusaccepted
contextevery currency change must be explainable
decisionappend-only ledger + idempotency keys
alsotransactional outbox for side effects
tradeoffmore rows, zero silent drift
step 02 · artifact
ops-agent.flowflow
read26 tools · no DB, no SSH, no SQL
write3 capped tools · daily quota
propose11 tools → admin_approvals
approvehuman, never the requester
auditevery action, with a reason
step 03 · artifact
ra2-bench · 64 playersbench
tick p501.6 ms
tick worst8.7 ms · none > 15 ms
bandwidth0.8 KB/s per player
dropped0
paritynative ≡ wasm to 1e-9
step 04 · artifact

step 01 / 04

Systems I designed, built and measured.

Each one is written up the way I would present it to a client: the constraint that made it hard, the architecture that answered it, and the numbers that prove it held.

xom-nong-trai / load-testspec
4,000CCU on 4 shards — all on one small 4-vCPU box
box4 vCPU · 8 GB, shared by Node, Postgres, Redis, nginx
server4 Node shards · ~1 GB each · nginx least_conn
postgres16 · 2.8 GB · max_connections 100 · pool 10 / shard
redis7 · 819 MB · AOF on
limitCPU-bound on the shared box; Postgres is the largest consumer
queriesPostgres queries per request 5.0 → 2.9
moneyappend-only ledger · idempotency keys · outbox
Personal project · private repository2026

012026Founder · built end to end

Xóm Nông Trại

Server-authoritative social platform with a real economy

The first product I took from a blank repo to live production across every discipline — 10,500+ players in its first month. A social farming game whose hardest problems are financial: every coin is money-like, players act concurrently, and live-ops staff need power without risk.

The constraint

Concurrent players, a currency that must never drift, and operators (human and AI) who need to change prices and issue refunds safely.

Architecture

  • Fastify + WebSocket server is the single source of truth; a shared rules package lets the client predict while the server decides.
  • Append-only ledger for every gold/xu change, idempotency key on every mutation, transactional outbox for quests, analytics and notifications.
  • Versioned config bundles with sha256 checksums — running jobs keep the version they started with.
  • Ops MCP: AI agents read through the admin API only and can only file proposals; refunds above a threshold need a second, different approver.

Evidence

  • 10,500+players in the first month after launch (Oct 2026)
  • 23 msAPI p95 in production at ~4.1M requests a day
  • 4,000CCU on 4 shards, one 4-vCPU box shared with Postgres + Redis
  • −82%draw calls on a full farm: 2,101 → 375
  • TypeScript
  • Fastify
  • PostgreSQL 16
  • Redis · BullMQ
  • Three.js
  • MCP
  • OpenTelemetry
Personal project · private repositoryPlay it live · nongtrai.jackle.dev

blank repo → live production

Zero to live — every discipline, one owner

No studio, no outsourced art or audio. I wrote the design, generated the world, made the sound, built the code, the release process and the operations that run it today.

  1. 01 Gameplay

    63

    design documents

    Core loop, Level-200 progression, quests, markets and social visits — balanced with 38 Python economy simulations before a config ships.

  2. 02 3D & art

    62

    GLB models from 92 Blender scripts

    A procedural low-poly kit — houses, crops, ten dog breeds — plus a style-locked sprite pipeline for 1,240 PNG assets.

  3. 03 Music & sound

    20

    audio cues synthesised in code

    Water, rod and wind modelled physically (Minnaert bubbles, Karplus-Strong); pentatonic score via MIDI → FluidSynth, seeded for identical renders. Background music composed with Gemini.

  4. 04 Build

    7

    packages, one TypeScript monorepo

    Three.js client, Fastify server, shared rules, admin console, ops MCP, playtest bots and a Capacitor iOS/Android shell.

  5. 05 Process

    913 + 171

    test files + Playwright e2e specs

    PR → CI → staging; bots play it on a 30× game clock, then the exact image they passed is promoted to production with auto-rollback.

  6. 06 Operations

    Daily

    AI ops review, human approval

    An ops agent reads metrics over MCP every morning and files proposals — fraud rings, price anomalies, balance changes. I approve or reject on Telegram. Nothing changes without a person.

3,550+ commits of my own · AI coding agents worked to my specs, tests and review gates

Performance · browser to database

Optimised to the bone — and measured at every layer

2,101 → 375draw calls on a fully built Level-200 farm. Each change below landed with its before/after number.
draw calls
Static parts merged by vertex colour; animals, dogs and swaying plants batched into BatchedMesh — 1,258 → 286 close up, 2,101 → 375 zoomed out.
gpu
Measured what screen-space ambient occlusion really cost, then made it opt-in: 9.0 → 5.2 ms of GPU time per frame.
download
3D kit 32 → 9.5 MB with meshopt + brotli, texture atlas 16.8 → 3.7 MB as WebP, texture memory 64 → 37 MB.
old devices
Devices graded A–D; quality tiers follow frame time, Low Power mode and battery. 2018 iPad 6: 36 → 50 fps.
server
p50 4.9 ms · p95 22.8 ms · p99 44.7 ms over 4.1M production API requests in one day, zero 5xx.

Built for the phones players actually own, not the laptop it was written on.

AI in production

The agent proposes. A person decides.

50+proposals filed by the ops agent and decided by a human — 48 of the latest 50 approved.
reads
Metrics, the ledger and player reports over MCP — through the admin API only, never the database.
fraud
Found an 8-account ring rigging a minigame leaderboard and alt accounts buying at 2–5× the median price to move gold; proposed the bans and clawbacks.
evidence
Every proposal carries its reason, a before-state snapshot and, for balance changes, a 7-day impact estimate: "102 of 807 steals would be blocked".
guardrails
40 tools, 11 of them proposal-only. Small writes are capped per day; refunds above a threshold need a second, different approver.
report
A daily 08:00 Telegram report — active players, retention cohorts, economy flows, API p50/p95/p99, threshold alerts — with one-tap approve or reject.

The shape of a claims agent: the AI reads and builds the case, a person signs off, the audit trail keeps both.

Cortex Hub dashboard showing projects, memory and agent sessions
Cortex Hub · production UI2026

022026Creator

Cortex Hub

One MCP endpoint for AI agent memory and code intelligence

Self-hosted platform that gives every AI coding agent on a team the same persistent memory, AST-aware code search, shared knowledge and quality gates.

The constraint

Agents forget everything between sessions and every tool reinvents context. Teams need one governed memory layer that works across clients, repos and machines.

Architecture

  • MCP gateway with API-key auth and telemetry in front of Qdrant vectors, an AST code graph and SQLite in WAL mode.
  • Knowledge store with temporal validity and supersession chains; recipes ranked 0.6 vector + 0.3 success rate + 0.1 recency.
  • LLM gateway with fallback chains, budgets and complexity routing; local embeddings via Ollama.
  • Conductor dispatches work across agents and machines with leases, file locks, review gates and escalation.

Evidence

  • 96.0%recall@5 · LongMemEval-S (500 q)
  • 97.8%recall@10 on the same set
  • ~10 mslocal embedding vs ~600 ms API
  • 890+commits · MIT licensed
  • TypeScript
  • MCP
  • Qdrant
  • SQLite
  • Ollama
  • Next.js
  • Docker
  • Cloudflare Tunnel
Open source · live at hub.jackle.devgithub.com/lktiep/cortex-hub
ra2-studio / benchspec
1.6 msserver tick p50 · 64 players · 0 dropped
coreRust · zero deps · fixed step
targetsnative + wasm32 · parity 1e-9
net30 Hz · delta snapshots 15/s
interestper-player camera + line of sight
renderWebGL2 · batched sprites
edgeWorkers · Durable Objects · Turnstile
20k unitsstep p50 3.5 ms · p99 8.3 ms
Personal project · private repository2026

032026Sole engineer

RA2 Studio

Deterministic real-time simulation, native and WebAssembly

A clean-room, web-native RTS engine: one Rust simulation core compiled to native and WASM, an authoritative multiplayer server and a WebGL2 client.

The constraint

Thousands of units, dozens of players, a browser client — and results that must be bit-identical everywhere so replays and anti-cheat can trust them.

Architecture

  • Dependency-free Rust sim core; a parity gate replays golden traces natively and in WASM and fails on drift beyond 1e-9.
  • Authoritative 30 Hz server streaming binary delta snapshots, filtered per player by camera and line of sight.
  • Room supervisor runs one process per match and restarts crashed rooms with backoff.
  • Edge security on Cloudflare Workers: Turnstile, single-use IP-bound HMAC join tickets, per-IP caps, pseudonymous identity.

Evidence

  • 1.6 mstick p50 · 64 players, 718 units
  • 0.8 KB/sbandwidth per player · 0 dropped
  • 26 → 4.4 msclient frame p99 after one fix
  • 60 fps64 × 500 units, frame p99 2.7 ms
  • Rust
  • WebAssembly
  • tokio
  • WebGL2
  • Cloudflare Workers
  • Durable Objects
  • WebRTC
Personal project · private repository
Shrimp Farm Platform dashboard with pond monitoring
Shrimp Farm Platform · production UI2024 — present

042024 — presentCTO & lead engineer

Shrimp Farm Platform

Multi-tenant AgriTech SaaS for commercial aquaculture

Farm management for operators running dozens of farms with thousands of ponds: real-time monitoring, water-quality analytics and AI camera detection.

The constraint

Field data arrives from sensors and phones in poor connectivity, while owners need one trustworthy view across every farm they run.

Architecture

  • Tenant-scoped data model across farms, ponds and crop cycles.
  • On-device TFLite detection with Gemini Vision as cloud fallback.
  • Real-time dashboards and alerting over the water-quality pipeline.

Evidence

  • Dozensfarms managed per operator
  • 1000sponds under monitoring
  • Next.js
  • Supabase
  • PostgreSQL
  • TFLite
  • Gemini
  • GCP
  • IoT
Kèo Thơm marketplace with live auctions
Kèo Thơm · production UI2025

052025Architect & full-stack engineer

Kèo Thơm

Real-time auction marketplace with escrowed wallets

Social marketplace for group-buy deals and live auctions, with an integrated wallet, frozen funds during bidding, seller verification and automated payouts.

The constraint

Bids land concurrently in the last seconds; money must be reserved, released or paid out exactly once.

Architecture

  • Wallet with frozen-funds holds so a bid is only valid while backed by reserved balance.
  • Real-time bid stream with server-side ordering; seller verification and automated payout jobs.
  • Next.js
  • Supabase
  • PostgreSQL
  • Cloudflare Pages
  • TypeScript
In productionkeothom.bid
  • Cortex bench harness

    Five-arm controlled experiment measuring what an AI agent still remembers on day two.

  • 3D Character Lab

    Image → Blender → WebGPU pipeline scored by silhouette IoU, SSIM and per-region ΔE.

  • game-dev skills

    Claude Code plugin with 7 skills distilled from shipping a live game: perf, content, monetisation guardrails.

  • CUGF Render Bench

    Local-GPU exploded-view renderer; WebP export cut each 96-frame set from 69 MB to 3.6 MB.

  • Spiral Studio

    Profile curve → spiral toolpath → G-code for five printer profiles, in the browser.

  • ZaloX

    CLI and REST API for Zalo messaging with webhooks, published on npm.

AI multiplies what I ship. Specs, gates and numbers keep it honest.

I work with AI coding agents every day. They write much of the code; I own the spec, the gates and the measurement. That is how one engineer shipped a live game across every discipline, and how I get productive in fields I had never worked in.

  1. 01

    Specify

    I write the contract before any agent touches code: boundaries, invariants, acceptance tests. The agent gets a spec, not a wish.

    brief.md · ADR · failing test

  2. 02

    Delegate

    Agents run in parallel through Cortex Hub — shared memory, leases, file locks, escalation — so several streams move at once without stepping on each other.

    conductor · leases · locks

  3. 03

    Gate

    Nothing merges on trust. CI, tests and my review stand between an agent and main; in production, agents can only file proposals a person approves.

    CI · review · proposal-only tools

  4. 04

    Measure

    Every claim ends in a number — draw calls, p95, recall@5. If it is not measured, it is not done, whoever wrote the code.

    bench · load test · telemetry

Eighteen years, engineer to CTO.

Global telecom clients, Australian enterprise SaaS, two CTO seats, and now independent platform work — hands on the code at every step.

  1. 2023 — now

    Principal Engineer & Architect

    Independent · AI & platform engineering

    Design and build production platforms end to end, from the first stakeholder conversation to the on-call runbook.

    • TypeScript
    • Node.js
    • Rust
    • PostgreSQL
    • Redis
    • GCP
    • Cloudflare
    • MCP
    • LLM
    • Built Xóm Nông Trại end to end — game design, 3D, audio, code, release and operations — now live: ledger-backed economy, idempotent mutations, outbox, AI ops agents limited to proposals that humans approve.
    • Created Cortex Hub, an open-source MCP platform for agent memory and code intelligence (96.0% recall@5 on LongMemEval-S).
    • Built a deterministic Rust/WASM simulation and authoritative multiplayer server, benchmarked to 1.6 ms tick p50 at 64 players.
    • Delivered a multi-tenant AgriTech platform and the Kèo Thơm real-time auction marketplace.
  2. 2022 — 2023

    Chief Technology Officer

    High-tech agriculture company

    Led architecture for a platform running thousands of hectares of production.

    • Microservices
    • Kubernetes
    • PostgreSQL
    • Big data
    • Real-time
    • Big-data pipelines processing millions of field data points for real-time monitoring and analytics.
    • Crop lifecycle, export-compliance traceability and live operational dashboards on distributed services.
    • Kept the system scalable and reliable under continuous data growth.
  3. 2021 — 2022

    Chief Technology Officer

    D2 Solutions

    Founded and led the team behind a cross-platform 3D multiplayer game with a Web3 economy.

    • C++
    • Unity
    • gRPC
    • Kafka
    • Redis
    • Kubernetes
    • AWS
    • GCP
    • Distributed real-time backend built to scale to millions of users with elastic micro-clusters of ~20 players per room.
    • Custom low-latency TCP networking layer that removed the dependency on Photon.
    • gRPC, Kafka and Redis on Kubernetes (HPA/VPA) with CI/CD and a full observability stack.
    • Auditable on-chain / off-chain token economy.
  4. 2018 — 2021

    Technical Architect

    Enterprise platforms

    Architecture for ERP, logistics and e-commerce platforms serving 100K+ users.

    • Microservices
    • Kubernetes
    • PostgreSQL
    • AWS
    • CI/CD
    • Microservices on Kubernetes with PostgreSQL and CI/CD on AWS.
  5. 2016 — 2018

    Senior Software Engineer

    Ansarada

    Enterprise SaaS for deal management and virtual data rooms.

    • .NET
    • Microservices
    • SQL Server
    • TDD
    • CI/CD
    • Engineered inside the most disciplined microservices architecture I have worked in: small services, explicit contracts, CI/CD on every change.
    • Test-driven development as the norm; kept unit-test coverage above 90% on the code I owned.
    • Tuned SQL Server queries and indexes on performance-critical paths — work singled out for praise by the CTO.
    • Moved from outsourcing to product engineering: features measured by customer value.
  6. 2008 — 2016

    Technical Architect · Senior Engineer

    TMA Solutions

    Grew from engineer to Technical Architect for the DG1 division (~300 engineers), serving global telecom clients.

    • C/C++
    • Telecom
    • Linux
    • Solaris
    • Multi-threading
    • Owned architecture, technical direction and mentoring across the division.
    • High-performance multi-threaded C/C++ on Windows, Linux and Solaris; led Windows → Linux and 32 → 64-bit ports down to drivers.
    • On-site with customers in the USA (twice, six months each), the UK and Australia (twice).
    • Outsourced delivery for banking and healthcare clients (client names under NDA).

Breadth to design the whole system, depth to debug any layer of it.

01Architecture

  • Distributed systems
  • Microservices & event-driven
  • Real-time & multiplayer
  • Multi-tenant SaaS
  • C4 · ADRs · data contracts
  • Security & threat modelling

02AI systems

  • LLM integration (OpenAI, Gemini, Claude)
  • OCR & document pipelines
  • Agents & MCP servers
  • RAG & vector search
  • Human-in-the-loop approval
  • Evaluation & benchmarks

03Cloud & platform

  • Google Cloud
  • AWS
  • Kubernetes (HPA/VPA)
  • Cloudflare Workers
  • CI/CD & IaC
  • Observability · OpenTelemetry

04Backend & data

  • Node.js · NestJS · Fastify
  • Rust · Go · C/C++ · .NET
  • PostgreSQL · MySQL
  • Redis · Kafka · BullMQ
  • gRPC · REST · WebSocket
  • Ledgers, idempotency, outbox

05Leadership & clients

  • Pre-sales & solution workshops
  • Exec & client presentations
  • Teams up to ~300 engineers
  • Delivery across Product · QA · DevOps
  • Code review & standards
  • Mentoring
InsuranceBankingHealthcareTelecomEnterprise SaaSGaming & real-timeWeb3AgriTech & IoTAI developer tooling

Outside the day job.

  • 01

    Winner, CodeProject Visual Studio Extension Contest

    Developer tooling

  • 02

    First Vietnamese to solve 300+ Project Euler problems

    Mathematics & algorithms

  • 03

    Products reaching 100,000+ users and downloads

    Consumer scale

  • 04

    Reverse-engineered a 3D online game client at assembly level

    Protocol reconstruction

Have a platform that has to be right?

I'm looking for a Solutions Architect or CTO role where I own the architecture end to end — in the room with clients, and in the code with the team. Happy to walk through any of the systems above in depth.

roles
Solutions Architect · CTO · Head of Engineering
work
On-site, hybrid or remote