Shipping generative AI into products at YouTube scale — hands-on, and honest about the ramp.
This is an AI/ML engineering seat on YouTube Knowledge — write and test product code, design and launch software, and ship solutions in a specialized ML area on top of YouTube-scale infrastructure. Fauzul brings the harder-to-teach half in full: a decade of shipping, testing and launching production software across nine geographies (TypeScript/Node, Go and Python), four years of that as a hands-on technical lead who set the code standards and architecture bar himself, and applied AI/ML in production — , a six-agent A2A mesh on Gemini 2.5 Flash Lite with Google Search grounding (his Google × Kaggle 5-Day AI Agents Intensive capstone); , an autonomous ReAct + evals pipeline whose dual-host dialogue is synthesised through Google's Gemini TTS multi-speaker mode (speech-style modifiers, interruption, temporal continuity, 16-bit PCM @ 24kHz stitched via ffmpeg) — speech/audio work directly adjacent to this team's domain; and this very site's chat agent running Google's on-device Gemma 4 E2B in-browser (WebGPU + MediaPipe). He'll be candid where the JD reaches past him: his applied-ML and model-deployment/evaluation experience is real, but three years of deep ML-training infrastructure and a single specialized ML field at hyperscale are an honest ramp, not a claim — and he closes ramps in public, the way the 912-day French streak proves. For a Paris seat, the French habit isn't decoration: it's evidence of exactly how he learns. He builds accessible technologies by default (WCAG 2 across five client organisations), which the team explicitly values.
Applied AI/ML, shipped — speech/audio, agents and on-device
I'll be exact about where I'm strong and where I'd ramp. The software and applied-AI craft is a decade deep; the YouTube-scale ML-infrastructure depth is the honest ramp — and I close ramps in public.
Speech synthesis through Gemini TTS — and a six-agent Gemini mesh
Two proofs in the team's adjacent space. : an autonomous ReAct + evals pipeline that synthesises dual-host dialogue through Google's Gemini TTS multi-speaker mode — speech-style modifiers, interruption, temporal and contextual continuity — stitched from 16-bit PCM @ 24kHz via ffmpeg. : six parallel country agents as FastAPI microservices on Gemini 2.5 Flash Lite, fed by an accuracy-critical sequential parser and synthesised by a strategic advisor, grounded against live Google Search at evaluation time — his Google × Kaggle 5-Day AI Agents Intensive capstone. Around both: a decade of shipping, testing and launching production software, and accessible-by-default delivery (WCAG 2 across five organisations).
: speech/audio synthesis via Google Gemini TTS multi-speaker (style modifiers, interruption, temporal continuity, 16-bit PCM @ 24kHz via ffmpeg) on a ReAct + evals pipeline — Milan AI Week 2026, results pending.
: A2A hub-and-spoke multi-agent architecture on Gemini 2.5 Flash Lite with Google Search grounding — solo build, MIT-licensed, submitted as the Google × Kaggle capstone.
On-device Gemma 4 E2B running this site's chat agent in-browser (WebGPU + MediaPipe, fully private) — model deployment at the edge, no server round-trip.
Data readiness done the hard way: a HIPAA-compliant patient-data pipeline at (2,212 new patients · 9,650 schedules in 2026), with CI/CD and observability built in.
Applied AI/ML — speech/audio, agents, evals, on-device
NewScriber on Google Gemini TTS multi-speaker (style modifiers, interruption, temporal continuity, 16-bit PCM @ 24kHz via ffmpeg) · VisaPros six-agent A2A mesh on Gemini 2.5 Flash Lite + Google Search grounding (Google × Kaggle capstone) · on-device Gemma 4 E2B in-browser via WebGPU + MediaPipe · explicit evals loops · Claude CMS in production.
Production software — shipped, tested, launched
A decade across nine geographies in TypeScript/Node, Go and Python; CI/CD and observability built himself (delivered 4x faster time-to-market (75% reduction in cycle time)); component libraries, NX monorepos and an OIDC/RBAC backbone; WCAG 2 accessibility across five organisations.
Data readiness, deployment & debugging
HIPAA-grade data pipelines (EKAGRA — 2,212 new patients · 9,650 schedules in 2026), media transcoding queues (AWS SQS/SNS) with low-latency audio/video sync, model deployment to Google Cloud Run and on-device — the unglamorous half of ML infrastructure.
Qualification by qualification
Mapped point-by-point to the posting's minimum and preferred qualifications — verbatim, with a verdict and evidence for each. The deep-ML-infrastructure and single-ML-field minimums are marked as honest ramps, with the bridge spelled out rather than dressed up.
Minimum qualifications
Bachelor's degree or equivalent practical experience.
Field of study: Computer Science & Engineering at North South University — then the equivalent-practical-experience route the posting explicitly allows: a UN internship (2016–17), co-founding in 2017, and nine-plus years of shipped production systems since.
5 years of experience with software development in one or more programming languages.
A decade (2017 → present) of production software in multiple languages — TypeScript/Node and Go as daily drivers, Python in agentic and data work ( is Python + FastAPI end to end). Shipped across health-tech, fintech, hospitality, music-tech and export, on Google Cloud Run, AWS and Cloudflare Workers.
3 years of experience testing, maintaining, or launching software products, and 1 year of experience with software design and architecture.
Far past the bar on both halves: nine-plus years launching and maintaining live products for paying clients, with testing and CI/CD pipelines he builds himself (GitHub Actions, delivered 4x faster time-to-market (75% reduction in cycle time)). On design and architecture, four years as a hands-on technical lead — component libraries, NX monorepos, multi-tenant architecture and an OIDC/RBAC backbone every product ships on.
3 years of experience with one or more of the following: Speech/audio (e.g., technology duplicating and responding to the human voice), reinforcement learning (e.g., sequential decision making), ML infrastructure, or specialization in another ML field.
The closest and most genuine match is speech/audio: synthesises dual-host dialogue through Google's Gemini TTS multi-speaker mode — speech-style modifiers, interruption, temporal and contextual continuity — stitched from 16-bit PCM @ 24kHz via ffmpeg, an autonomous ReAct + evals pipeline. Sequential decision-making shows up too: decomposes a task across a sequential accuracy-critical parser → six parallel agents → a synthesis advisor. What he won't claim is three years narrowly inside one of these fields at research depth.
He'd enter as a strong applied engineer in this team's adjacent space — speech/audio synthesis and agentic, evals-driven pipelines — and specialize into YouTube Knowledge's chosen ML area on the job. The craft that transfers (data pipelines, evaluation harnesses, deployment, debugging) is exactly what the role lists; the field-specific depth is the ramp, and he'd close it the way he closes every gap — in running code, fast and in public.
3 years of experience with ML infrastructure (e.g., model deployment, model evaluation, optimization, data processing, debugging).
Real on several of the listed pieces: model deployment (LLM-backed services to Google Cloud Run, plus an on-device Gemma 4 E2B model deployed in-browser via WebGPU + MediaPipe on this very site), model evaluation (explicit evals loops in and ), and data processing (HIPAA-grade patient-data pipelines at — 2,212 new patients and 9,650 schedules in 2026). The honest gap is tenure inside dedicated model-training-and-optimization infrastructure at YouTube scale.
Deployment, evaluation, data processing and debugging he does today; the missing piece is large-scale training/serving infrastructure and model optimization as a sustained specialty. He'd ramp it on Google's own platform — the fastest place on earth to learn it — and brings production discipline (CI/CD, observability, governance) that ML-infra work usually lacks.
Preferred qualifications
Master's degree or PhD in Computer Science, or a related technical field.
No graduate degree — saying so plainly rather than dressing it up.
The applied equivalent is public and testable: the Google × Kaggle 5-Day AI Agents Intensive capstone (), speech/audio synthesis through Gemini TTS, on-device Gemma, evals pipelines, and a decade of production systems. The preferred qualification proxies for depth — he's happy to be tested on ReAct loops, evaluation design, state management and tool-calling in running, open-source code.
5 years of experience with data structures and algorithms.
A decade of production engineering with data structures and algorithms as daily work — concurrency and queueing (AWS SQS/SNS, low-latency audio/video sync), graph-shaped multi-agent orchestration, and the accuracy-critical parsing and synthesis pipeline inside .
1 year of experience in a technical leadership role.
Three-plus years as a technical leader — Software Engineering Manager (Dec 2022 – Sept 2024), then CIO (Sept 2024 – present) — setting code standards, owning architecture, and mentoring deeply (two alumni now lead at organisations in Norway). Comes in able to lead by example as a senior IC and lift the engineers around him.
Experience developing accessible technologies.
Accessibility is a default, not an afterthought: came up as a UX Engineer and shipped WCAG 2 compliance across five client organisations. The team states it 'welcomes people with disabilities' and that everyone deserves a voice — building for that is already how he works.
Strengths & honest gaps
A decade of shipping, testing and launching production software — still hands-on
Nine-plus years writing, testing and launching live products across nine geographies in TypeScript/Node, Go and Python, with CI/CD and observability he builds himself (delivered 4x faster time-to-market (75% reduction in cycle time)). Came up as a UX Engineer, grew through Principal Staff SWE into management, and never stopped shipping — the kind of senior IC who writes the code and reviews the hard part.
Applied AI/ML in production — speech/audio, agentic systems, evals, on-device
Directly adjacent to YouTube Knowledge's domain: synthesises dual-host dialogue through Google's Gemini TTS multi-speaker mode (speech-style modifiers, interruption, temporal continuity, 16-bit PCM @ 24kHz via ffmpeg) on a ReAct + evals pipeline; is a six-agent A2A mesh on Gemini 2.5 Flash Lite with Google Search grounding, his Google × Kaggle Agents Intensive capstone; and this site's chat agent runs Google's on-device Gemma 4 E2B in-browser via WebGPU + MediaPipe.
Data readiness, pipelines & debugging at scale
The unglamorous half of ML infrastructure is where his decade lives: a HIPAA-compliant patient-data pipeline and EHR integration at (2,212 new patients, 9,650 schedules in 2026), media transcoding queues (AWS SQS/SNS) with low-latency audio/video sync, and a Claude-powered natural-language CMS running in a 26-year-old export house's production stack. Data processing, deployment and debugging as everyday work.
Accessible technologies, by habit
A UX Engineer's instinct carried into a decade of delivery: WCAG 2 compliance shipped across five client organisations, and an explainable, context-engineered AI interface on this very page. The role explicitly values developing accessible technologies — and a team built on 'everyone deserves a voice' — and that's a default in how he builds, not a line item.
Genuinely building toward Paris
A 912-day unbroken daily French streak, A2 and climbing, sharing his own learning in public every week as @FrenchwithFauzul — that's not a language flex, it's evidence of exactly how he closes any gap: relentless, daily, in the open. The same discipline he'd point at this team's ML field and at YouTube-scale infrastructure — and a genuine, intentional pull toward Paris.
Three years of deep ML-training infrastructure and a single specialized ML field at hyperscale.
First quartersHis applied ML is real — speech/audio synthesis, multi-agent decomposition, evals, model deployment, data processing — but a sustained specialty inside model-training/optimization infrastructure at YouTube scale is the honest ramp. He closes it on Google's own platform with production discipline most ML-infra work lacks (CI/CD, observability, governance), and ramps visibly — the same way the 912-day French streak proves he learns.
Systems at YouTube scale — billions of users, videos and watch-time hours.
Weeks, not quartersHe's a cloud-agnostic architect who designs and ships large-scale systems (multi-tenant platforms, queue-driven media pipelines, distributed remote-first delivery) — but firm-and-enterprise scale, not hyperscale. The distributed-systems and data-readiness instincts transfer directly; what's new is the order of magnitude, and he'd learn the YouTube-scale primitives before changing anything.
Candid answers
The questions a YouTube hiring manager should ask — answered candidly: why a founder-CIO wants an IC seat, where the ML-infrastructure experience actually is versus where it isn't, Claude versus Gemini, the missing graduate degree, YouTube-scale systems, and whether Paris is real.
You co-founded ELO and serve as its CIO — why step into an individual-contributor Senior AI/ML SWE seat?
Because the part of the work I love most is building, and this role is squarely that — write and test product code, design and launch software, ship solutions in a specialized ML area at a scale I can't touch anywhere else. I came up as an engineer and never stopped shipping; the Claude CMS, and are recent personal builds, not delegated ones. I've done the management arc and I'm good at it, but I want to be hands-on inside YouTube-scale ML, learning the field-specific depth from the people who set the bar — and lifting the engineers around me as a senior IC while I do it.
The role wants three years of ML-infrastructure experience — model deployment, evaluation, optimization, data processing. Be honest about where you actually are.
Honestly: I have real experience on deployment, evaluation, data processing and debugging, and an honest ramp on training-infrastructure and optimization as a sustained specialty. Concretely real — I deploy LLM-backed services to Google Cloud Run, I deployed an on-device Gemma 4 E2B model in-browser via WebGPU and MediaPipe, I run explicit evals loops in and , and I've built HIPAA-grade data pipelines at . What I won't claim is three years living inside large-scale model-training and optimization infrastructure. That's the piece I'd build on Google's platform — and I'd bring production discipline (CI/CD, observability, governance) that a lot of ML-infra work is missing. A team's trust comes from being exact about this, not from retrofitting a résumé.
The third minimum asks for a specialized ML field — speech/audio, RL, or another. Where's yours?
Speech/audio is the most genuine, and it's directly adjacent to YouTube: synthesises dual-host dialogue through Google's Gemini TTS multi-speaker mode — speech-style modifiers, interruption, temporal and contextual continuity — stitched from 16-bit PCM at 24kHz via ffmpeg, on an autonomous ReAct + evals pipeline. There's sequential decision-making in ' parser-to-agents-to-advisor decomposition too. What I won't overstate is research-depth tenure in a single field. I'd come in as a strong applied engineer in the team's adjacent space and specialize into YouTube Knowledge's chosen area on the job — the data, evals and deployment craft transfers; the field depth is the ramp.
Your production AI is Claude-led. This is a Gemini and Google-stack shop, at YouTube.
I'll be honest rather than retrofit: the AI I've shipped into client production has been Claude-led. But Gemini is genuinely second in my stack, not a gesture — runs six parallel agents on Gemini 2.5 Flash Lite with Google Search grounding and was my Google × Kaggle Agents Intensive capstone; renders its podcasts through Google's Gemini TTS multi-speaker mode; this site's chat agent runs Google's on-device Gemma 4 E2B; my deploys already run on Google Cloud Run; I use Google Antigravity as my core dev harness and shipped when its 2.0 release broke my workflow. The craft — evals, data pipelines, deployment, debugging — is model-agnostic, and the platform specifics I'd ramp visibly.
The posting prefers a Master's or PhD — you have neither.
Correct, and I won't pretend otherwise. My field of study was Computer Science & Engineering at North South University, and I took the shipped-artifact route from there: the Google × Kaggle Agents Intensive capstone, speech/audio synthesis through Gemini TTS, on-device Gemma, evals pipelines, nine-plus years of production systems. The preferred qualification proxies for depth in ML-system design — evaluation, state management, tool-calling, sequential decomposition — and that's exactly what I can demonstrate in running, open-source code rather than a thesis. I'm happy to be tested on it at whatever depth the panel wants.
You've shipped at firm and enterprise scale, not YouTube scale — billions of users, videos and watch hours. Why trust you here?
Fair, and I won't overstate it. What transfers directly is the systems instinct: I'm a cloud-agnostic architect who's built multi-tenant platforms, queue-driven media pipelines with low-latency audio/video sync, and distributed delivery across nine geographies. What's new is the order of magnitude — hyperscale data and serving — and I'd treat that with respect: learn the YouTube-scale primitives and the existing infrastructure before changing anything. The data-readiness, evaluation and debugging discipline I bring is exactly what holds up when scale punishes sloppiness.
How do you stay hands-on and technical as a senior IC when you've spent years as a manager?
Player-coach is the only way I've ever worked, and this seat lets me lean into the player half. I never stopped shipping through management — , , the Claude CMS, the on-device Gemma chat agent and are all recent personal builds. As a senior IC I'd write and test product code, do design and code reviews, set standards by example, and mentor the engineers around me without carrying a manager's org load. The leadership multiplies the building here; it doesn't replace it.
The role is in Paris, and many of your prospective teammates and users are French-speaking. Are you genuinely set up, and what does your French habit signal?
Straight answer: I'm at CEFR A2 and climbing, not fluent — so for anything that genuinely needs fluent French today I'd pair rather than fake it; respect beats bravado. But the habit is the real signal. I've kept a 912-day unbroken daily French streak and I share my own learning in public every week as @FrenchwithFauzul — that's evidence of exactly how I close a gap: relentless, daily, in the open, with no audience to perform for. It's the same discipline I'd point at this team's ML field and at YouTube-scale infrastructure. And I'm genuinely intentional about Paris — ready to land on CET from day one and keep getting measurably better in the language while I'm there.
The stack & the builds
The stack behind the verdicts — applied AI/ML in shipped code (speech/audio, multi-agent, evals, on-device), the Google-stack ties, and the builds that prove it.
Agentic news scraper & editorial audio briefing network. Scrapes tech and business news via Firecrawl, curates & scripts dual-host dialogues with Kimi via OpenRouter, and renders high-fidelity multilingual voices via Gemini TTS on Azure.
Submitted as the capstone for the Google × Kaggle 5-Day AI Agents Intensive (2026). VisaPros is an agentic visa eligibility advisor utilizing a Google ADK agent-to-agent (A2A) hub-and-spoke architecture — parsing documents once and running six parallel country agents as FastAPI microservices to evaluate fit across destinations.
Strategic health-tech partnership and equity investment in Bangladesh's most advanced clinical practice for wound care, diabetes, and nephrology. Spearheaded high-utility, HIPAA-compliant EHR clinical dashboards and scaled digital patient advocacy.
Editorial export site for Global Jute Trading Ltd — a Dhaka-based sourcing & exporting house with 26 years as a company and 35 years of jute trading under its principals. Built to support their push into deeper international markets, with a particular focus on Canada.
Open-source Windows GUI + CLI that links Windows Host directories straight to the WSL 2 Linux filesystem via native directory symlinks. Born the week Google's Antigravity IDE 2.0 broke the WSL Remote workflow — built to restore zero-latency, native-Linux-speed cross-platform development on a Windows desktop.
Everyone deserves a voice. Let's build at that scale.
Paris is where I want to do this work next — and I'm genuinely intentional about it, ready to start on CET from day one. The French side is personal: a 912-day daily streak, sharing my own learning in public as @FrenchwithFauzul. I'd arrive already living the language, not just visiting it.
See my Paris relocation plan →