Vadi Taslim

AI Systems Engineer · Senior Software Engineer | Sydney, Australia

Get in touch

Experience

August 2021 - Present

Millions of users · 30+ countries

Senior Fullstack Engineer at Vervio

▹Cut JS bundle in half—Core Web Vitals improved noticeably
▹Fixed SEO breakage that had been stuck across international sites
▹Stopped the platform from randomly crashing under load

Show 5 more...

▹Full-stack work across millions of users in 30+ countries—web, CMS, search, the whole stack.
▹Cut JS bundle by 50% with code splitting and i18n optimization.
▹Built async rendering to stop the UI from freezing while waiting for the backend.
▹Ripped out the legacy CMS and moved to Next.js. Freed up SEO that had been broken for months across non-US sites.
▹Set up ESLint, Prettier, pre-commit hooks. Reviewed code until the team actually remembered to run them.

Next.js·React·TypeScript·Node.js·Python·AEM

August 2020 - Present

100k+ Discord bot users

Senior Frontend Engineer at Toptal

▹Discord bot hit 100k+ active members
▹SSR retrofit cut initial load time by 40% and fixed SEO
▹Built email templates that actually work with screen readers and system themes

Show 4 more...

▹Built a Discord bot from scratch. Got 100k+ active users without it falling over—Node, Redis, Postgres.
▹Bolted SSR onto an existing React/Redux app without breaking CSS Modules, routing, or state.
▹Set up CI/CD with GitHub Actions and Docker to S3.
▹Owned the frontend across multiple projects. Reviewed code every day until people stopped making obvious mistakes.

React·TypeScript·Redux·Node.js·Redis·PostgreSQL

November 2018 - July 2020

Frontend Engineer at Wallethub

▹Automated CVE detection instead of manual hunting
▹Fixed security holes from bug bounty reports
▹Built the test suite—zero regressions after it went in

Show 4 more...

▹Picked up OWASP ZAP in week one and owned security automation from there.
▹Automated vulnerability scanning for the OWASP Top 10—replaced the manual review process.
▹Triaged bug bounty reports from Bugcrowd and HackerOne. Fixed CVEs in PHP and Angular.
▹Shipped the first actual test suite using Mocha. Stopped regressions from going to prod.

Angular·TypeScript·Node.js·PHP·OWASP ZAP·Redis

Earlier experience (2012–2019)

Frontend Engineer (3D/WebXR)atAwe2017–2019

▹Built the 3D engine every client project relied on
▹Killed memory leaks that were crashing sessions
▹Kept WebGL/VR/AR working on every browser

Show 4 more...

▹Built the 3D rendering engine all client projects used.
▹Learned the math—matrices, trig, linear algebra—built it all from scratch.
▹Shipped Three.js plugins and kept WebGL/XR working across browsers.
▹Found and fixed memory leaks that were nuking long-running sessions.

Three.js·WebGL·WebXR·JavaScript·Gamepad API

Frontend EngineeratDominopos2016–2017

▹Got the team writing tests with Selenium/Nightwatch
▹Placed 2nd in the hackathon
▹Built the first test suite for the whole company using Selenium and Nightwatch.

Show 1 more...

▹Won 2nd place at the hackathon—built an Angular/Node app in 48 hours.

Angular·TypeScript·Laravel·MongoDB·Elasticsearch

Fullstack EngineeratLCA2015–2016

▹Zero regressions to prod because tests were there from the start
▹Built backend features with tests baked in from day one—not bolted on later.
▹Ripped out dead code and measured how much faster things got.

PHP·Symfony·Python·PHPUnit·Selenium

Software DeveloperatPowergen2012–2015

▹Shipped a biometric ID system that went into enterprise
▹Built a facial recognition and fingerprint ID desktop app.
▹Integrated DigitalPersona and Neurotech biometric SDKs.

Show 1 more...

▹Debugged production issues on-site—hardware integration and matching bugs.

C#·.NET·MSSQL·DigitalPersona·Neurotech

Projects

I build what I can't stop thinking about — usually because I want the thing to exist and nobody else has built it yet.

Vargos.ai

Personal AI agents are cool but useless. They can't access your tools, and they disappear after each message.

An MCP server that runs AI agents—gives them persistent tool access, channels (WhatsApp, Telegram), and the ability to actually do stuff on your server.

Learnings:

▹LLM orchestration at system level
▹Vector search for semantic function discovery

TypeScriptMCPPostgreSQLpgvectorSQLiteWebSocketPlaywright

Chans.ai

Slapping voice onto an AI agent means cobbling together STT, TTS, and WebRTC yourself. That sucks.

Drop real-time voice into any LLM. It's just an API—STT, TTS, voice chat, all connected to your model.

Learnings:

▹Real-time audio streaming with LiveKit and WebRTC
▹Hybrid RAG search (vector + BM25) with line-level citations

PythonFastAPILiveKitNext.jspgvector

Archived

Ingra.ai

A place to dump functions and let LLMs call them as tools.

Tic Tac Toe

A tic-tac-toe AI you can't beat because it's using minimax.

URL Shortener

A quick link shortener so I don't have to copy/paste long URLs between my phone and desktop.

Writing

View all

Writing about the decisions behind the systems — why something was built that way, and what went wrong first.

ai·Mar 2026·6 min

Nemotron 3 Super vs Qwen 3.5: Tested on a Real Agentic Workflow

One task, no synthetic prompts: SSH into a homelab server and document the infrastructure autonomously. nemotron-3-super:120b made zero tool calls. qwen3.5 made forty-three and built a cross-linked wiki.

ai·Mar 2026·3 min

Where Ollama Stores Your Models (And How to Use Them Directly)

Ollama stores downloaded models as plain GGUF files. Here's how to find them, use them with llama.cpp tools, and what breaks when you try.

ai·Mar 2026·9 min

Getting Started with NVIDIA NeMo RL + LLM

A practical guide to teaching LLMs to learn from feedback using NVIDIA NeMo RL - no PhD required.

About

In Sydney, I started with game design, then switched to game design, and then changed to systems and infrastructure, over the years I've built a biometric identity platform, three-dimensional printing machine has a slightly biased direction, please allow me to correct and complete this translation task, to make sure it fully meets your requirements and keeps the natural flow of Thai expression: in Sydney, I began with game development, then moved to complete system and infrastructure Over the years I have used it, I have built a biomechanical identity platform, a three dimensional printer and a gaming currency system, which is used in decimal and decimal jobs. in game design I've been looking for ways to do things, not just destroy the basic way it works.

Right now I'm working from home using GPU passthrough GPUs for local AI, writing the OWASP ZAP rules myself because some of the configuration is problematic, and hosting most of what I rely on. 15 years later, I still learn a lot of new things from mistakes.

Technologies

Frontend

Fast, SEO-friendly interfaces that handle real users

TypeScriptPrimary language

React10+ years

Next.jsApp Router, SSR, ISR

Tailwind CSSDesign systems

Backend

APIs that don't fall over, data pipelines, automation

Node.jsProduction services

PythonAI/ML, automation

PostgreSQLPrimary datastore

RedisCaching, queues

Infrastructure

Hardware I own, operate, and repair

DockerContainer orchestration

ProxmoxVMs, LXCs, GPU passthrough

CloudflareDNS, tunnels, edge

LinuxDaily driver

AI & ML

Running models locally without selling my soul to the cloud

OllamaLocal LLM serving

llama.cppQuantized models

WhisperSpeech-to-text

Stable DiffusionImage generation

LiveKitReal-time voice AI

Infrastructure

Cloud abstractions hide how systems actually behave. I self-host to see what gets hidden.

Intentional boundaries: Compete with hyperscalers · Replace managed services at scale · Build infra without operational purpose

What self-hosting taught me

▹Every abstraction leaks; knowing when matters more than avoiding them
▹Cold starts, IO contention, and memory pressure define real performance
▹Don't pick GPU specs until you know your actual workload
▹VRAM is the real GPU bottleneck — not compute
▹The bug is almost always in the network. It just looks like your app.
▹Disk and memory constraints surface before CPU saturation
▹Backups are useless until restores are tested
▹Observability beats intuition
▹Infra that isn't automated will eventually drift

Mobile dev environment

Mobile terminal

Termux (Android)cloudflaredtmuxClaude CLI

Full development environment from a phone — tunneled into my server, persistent sessions, agentic coding with Claude.

AI agent via chat

WhatsAppVargosHome server

Message my AI agent on WhatsApp to execute tasks directly on my infrastructure.

Details

Machines

HYDRA

Primary Workstation · Windows 11 + WSL2

Intel Core i9-12900K

128GB DDR5 4400MHz

NVIDIA RTX 5080 (16GB)

2TB NVMe + 2x1TB NVMe

▹All-purpose development workstation
▹Local ML experimentation
▹GPU-bound workloads and tooling

HYDRA09

Primary Server Node · Proxmox VE

AMD Ryzen 9 7950X

192GB DDR5 4000MHz

NVIDIA RTX PRO 6000 Max-Q

2x2TB NVMe + 4TB NVMe + 2x18TB HDD RAID 1 ZFS-backed bulk storage

▹Virtualization (VMs & LXCs)
▹LLM and Stable Diffusion inference workloads
▹Self-hosted services (DNS, tunnels, media, databases)

Beelink SER9 Pro

Edge / Utility Node · Proxmox VE

AMD Ryzen 9-class mobile CPU

32GB DDR5

1TB NVMe

▹Edge services
▹Lightweight containers
▹Auxiliary workloads and testing

LEGION-S7

Mobile Control & Development Node · Ubuntu

AMD Ryzen 7 5800H

32GB DDR4

NVIDIA RTX 3060 (Laptop)

1TB NVMe

▹Primary mobile workstation for development

Networking

MikroTik RouterOS

▹Subnet and bridge isolation
▹NAT and firewall rule ownership
▹WireGuard VPN for secure remote access
▹Cloudflare-based DNS and ingress control

Platforms & Tooling

AI & ML Tooling

Local LLM runtimes (llama.cpp, Ollama)Model experimentation and benchmarkingGPU utilization and performance analysis

Virtualization & Containers

Proxmox VE (VMs and LXCs)PCIe device passthrough (GPU, storage, networking)DockerDocker ComposeContainerized services on Proxmox LXCs

Networking & Access

MikroTik RouterOSWireGuard VPNCloudflare DNS and Tunnel-based ingressReverse proxy and internal routing

Data & Storage

ZFS (datasets, snapshots)NFS mounts between nodesNAS-backed persistent volumesPostgreSQLRedisMariaDBVector databases (Qdrant)

Media & Real-Time

Plex media streamingPhotoPrism photo indexingLiveKit real-time audio/video

Observability & Ops

Service health monitoringLog inspection and debuggingBackup orchestration and restore testing

Get in touch

Tell me what you're building.

[email protected]

Developer Q&A reputation

Top 3% of vetted developers

GitHub contribution streak statistics for Vadi Taslim

Open source contribution streak