Dylan Ler, Senior AI Automation Architect at Product.ai
Senior AI Automation Architect

Dylan Ler

Senior AI Automation Architect · Product.ai

Designs, builds, and ships AI automation pipelines across Product.ai — turning repeatable manual work into grounded, human-reviewed systems. An engineer whose arc runs electrical engineering → data science → AI architecture, building production automation since before AI tooling went mainstream.

Runs several automation pipelines in production today — discovery, evaluation, generation, classification, content. Every one of them replaced something a person used to do by hand. The method is consistent: ship the simple version, watch what real data does, then build the next stage on the evidence.

LinkedIn → GitHub → Durian Data Lab (Medium) →
Role
Senior AI Automation Architect
Based
Los Angeles, CA
Focus
Automation pipelines · AI-as-leverage
Prior
Udacity · WebMD · Green Robin Ventures · UCLA EE
10yr
Building ML systems — since before AI tooling went mainstream
5+
Production programming languages in shipping code
5
Spoken languages — English, Chinese, Malay, Indonesian, Cantonese
2016
First open-source ML projects published
About

The pipeline engineer who turns repeatable manual work into grounded, human-reviewed automation — and has been building ML systems for a decade.

10yr
building ML systems
5+
production languages
5
spoken languages
2016
first ML projects published
The range

01 · Core discipline

Automation Pipeline Architecture

Replaces a repeatable task — discovery, evaluation, generation, classification — with a system that runs it. Several pipelines in production today, each one author’s work end to end.

02 · Core discipline

Applied Machine Learning

A decade of hands-on ML, from self-driving coursework and mentoring to production model pipelines. The fluency that wires models into real systems, not demos.

03 · Core discipline

Systems & Data Engineering

Electrical-engineering roots into full data pipelines — ingestion, enrichment, classification, serving. Five-plus production languages in shipping code.

04 · Sharpened in production

AI-as-Leverage

Composes grounded, human-reviewed automation instead of black-box autonomy. Each step has a named contract; a person ratifies anything that touches production.

05 · How he debugs

Population-First Diagnosis

When one case breaks, he audits the whole class before designing the fix. The solution falls out of the actual distribution, not the textbook example.

06 · Off the resume

Cross-Discipline Range

Engineering, data science, founding, and investing in one operator. Each chapter feeds the current role — pricing scope, decomposing work, thinking in systems.

Marquee stops

Electrical engineering → data science → founding → AI automation architect.

2025 → Now · Product.ai
Product.ai
Senior AI Automation Architect
Designs, builds, and ships the automation pipelines that convert manual hours into operator leverage — discovery, evaluation, generation, classification, and content systems, several of them in production. The pattern is consistent across all of them: grounded, human-reviewed automation with a named contract at every step, shipped fast and improved on real production evidence rather than on paper.
Pipelines in production
Before Product.ai
~2022 → 2025 · Demand.io / Product.ai
Data Scientist → AI Engineer → Architect
IC arc inside the company
Joined as a data scientist focused on AI insights and personalization, then grew through AI engineering into the company’s first Senior AI Automation Architect as the automation portfolio expanded from one system to many. The role grew with the substrate.
Role grew with the work
2021 → Present · Aki + Green Robin Ventures
Co-founder, Aki · Partner, Green Robin Ventures
Web3 founding + early-stage VC
Co-founded Aki, a Web3/NFT venture during the 2021–22 cycle, and invested in early-stage startups as a partner at Green Robin Ventures. The founder and investor muscle shows up today in how he prices scope and ramps systems.
Founder · Investor
2017 → 2022 · Udacity
Udacity
Self-Driving Car Nanodegree mentor · 5 years
Five years mentoring autonomous-systems engineers through Udacity’s Self-Driving Car Nanodegree — starting before machine learning became consensus. The same muscle now shows up as fluency wiring models into production pipelines.
Autonomous-systems mentor · 5 yrs
~2016 → 2020 · WebMD · Appable · bVentures
WebMD · Appable · bVentures
Data Scientist · Product Engineer · Director
Data Scientist at WebMD before the AI-in-healthcare boom, Product Engineer at Appable, and Director at bVentures. The path that established the IC-to-architect range visible today.
IC range across industries
2016 → 2017 · Durian Data Lab
Open-source ML experimentation
Personal projects · Medium @duriandatalab
Personal ML projects published starting 2016 — an RNN that generates Singlish, image classification of regional Asian foods, AI-art experiments (some published in HackerNoon). Same instinct as today: pick a small, real problem and build the thing that solves it.
First ML systems · 2016
2012 → 2016 · UCLA
UCLA
BS, Electrical Engineering · Bruin Entrepreneurs
Undergraduate EE — circuit analysis, systems & signals, early machine-learning coursework. Hardware projects ranged from a MYO-band sign-language interpreter to a medical-aid mobile app to an autonomous-vehicle project. Active in Bruin Entrepreneurs and the Malaysian Student Association.
UCLA EE
2011 · Sabah, Malaysia
Raleigh International (volunteer)
Community structure-building
Two months building community structures in rural Malaysia. The international, multilingual orientation that runs through his profile has roots earlier than the resume captures.
Service · pre-career
How he thinks

Ship first. Then look at what actually broke.

Dylan’s method has one half he says out loud and one half he just does.

The half he says out loud: ship the simple version first, then iterate on real data. He’d rather get version one into production — processing real records, real candidates, real signal — than spend another week speccing in the abstract. A first production run tells him more in one afternoon than a month of planning would.

The half he just does: when something breaks, he audits the whole class before designing the fix. One flagged duplicate becomes a read-only audit across hundreds of cases, sorted into real categories — and the fix design falls out of the actual distribution rather than the single example. The diagnostic that surfaces the whole class is cheaper than case-by-case whack-a-mole.

Your three-hour weekly chore is my next fifteen-minute skill. When a teammate is slowed down by a manual process, I’m already scanning for the next thing to automate. Dylan, on what he’s always looking for
Operating code

Six principles visible across the work.

01 Method

Ship version one, then iterate on real evidence.

A week of speccing produces less signal than the first afternoon of real data. Every system he’s shipped has improved faster once it was live than it did on paper.

“I’d rather get version one into production processing real records — then look at what actually broke or looked weird — than spend another week speccing in the abstract.”

02 AI posture

AI is a grounded skill runtime, not an autonomous agent.

Pipelines compose well-grounded steps with human approval gates rather than ungrounded autonomous loops. Each step has a named contract; the grounding is explicit.

“The output is better because each step has a named contract, the grounding is explicit, and a human ratifies anything that touches production.”

03 Diagnosis

Audit the population before designing the gate.

When one case is flagged, build the diagnostic that surfaces the whole class first. The fix design falls out of the real distribution, not the textbook case.

“Build the gate against the actual distribution, not the textbook case. The single-case guard would have shipped a whole false-positive class.”

04 Stewardship

Fix the class, not the instance.

One cleanup fixes today; the right workflow makes the next occurrence structurally impossible. Stewardship compounds when fixes go after classes, not instances.

“Stewardship compounds when fixes go after classes, not instances — one workflow makes a known failure pattern structurally impossible to recur.”

05 Working mode

A rough brief beats a fully-specced doc.

Tell him the outcome and let him propose the shape, and you’ll get more value than handing him a finished forty-item spec. Deep-work time is protected so the building actually happens.

“Give me a rough brief and let me propose the shape. Tell me the outcome and let me architect the pipeline.”

06 North star

Every system runs the same maturity arc.

Deterministic rules first, then a model on top with an audit trail, then feedback loops — with a human ratifying every model update. The arc is the architecture, not an aspiration.

“The loop is closed by humans, not by the pipeline marking its own work correct.”

Built in the open

A decade of open-source ML, published on GitHub.

The instinct behind the production work shows up on the open web too — a long public trail of machine-learning and automation experiments. A representative slice:

agentic-video-generator

GitHub

Multi-step video generation from a prompt.

An experiment in chaining generation steps into a working video pipeline — the same compose-grounded-steps instinct now visible in production tooling.

latent-planning

GitHub

Planning experiments in latent space.

Research-flavored exploration of how models plan — the kind of hands-on ML work that predates the current wave.

AI-auto-narrator

GitHub

Automated narration over generated media.

A small, real problem solved end to end — his recurring pattern: pick the concrete thing and build the system that does it.

digital-twin-hub

GitHub

A hub for digital-twin experiments.

Systems thinking applied to simulation and representation — EE roots showing through in how he frames a build.

ai-screen-studio

GitHub

Screen-based AI tooling experiments.

Tooling that wires models into a real interface — the through-line from personal projects to production automation.

Durian Data Lab

since 2016

The personal ML lab behind the GitHub trail.

The umbrella for years of open ML work — from a Singlish-generating RNN to regional-food image classification. The substrate has been there a decade.

Published thinking

Open-web machine-learning writing, since 2016.

Medium · Durian Data Lab 2016→

Durian Data Lab — the personal ML blog

Hands-on machine-learning writeups — RNN-generated Singlish, regional-food image classification, AI-art experiments. The public record of a decade of ML tinkering.
GitHub · @dylanler Ongoing

Open-source ML & automation experiments

A long public trail spanning agentic video generation, latent planning, digital twins, and screen-based AI tooling. The same instinct that ships the production pipelines.
What sets him apart

Combinations rare individually. Unusual to find in one operator.

Many production pipelines, one author

Discovery, evaluation, generation, classification, content — same engineer, same approval-gated pattern. Most teams run several pipelines with several engineers; he runs several as one.

The cross-discipline arc

EE → data scientist → self-driving mentor → founder → investor → automation architect. Each chapter feeds the current role: VC for pricing scope, mentoring for decomposing work, EE for systems thinking.

A constraint-respecting AI builder

Composes grounded automation with human approval gates rather than chasing black-box autonomy. The substance stays high-leverage; the framing stays honest about what AI should and shouldn’t decide alone.

Audit-first debugging

When one case breaks, he inspects hundreds before designing the fix. Build against the actual distribution, not the textbook example — the same shape across every cleanup.

A decade of ML before it was mainstream

Self-driving coursework, five years mentoring autonomous-systems engineers, and open-source ML projects since 2016. The fluency wiring models into systems is earned, not new.

Trilingual-plus engineer with a public ML trail

Five spoken languages, five-plus production programming languages, and open-source ML work on the web since 2016. The range is real and long-running.

Career history, open-source projects, and published writing on this page are publicly verifiable. See the record →
The through-line

Find the manual chore. Build the grounded system that replaces it. Keep humans on every production-state change. Compound the leverage one pipeline at a time.

Product.ai builds with operators like Dylan. See open roles →

Get in touch with Dylan See the team