Cansu Kaya, Data Supply Chain Architect / Analytics Engineer at Product.ai
Data Supply Chain Architect / Analytics Engineer

Cansu Kaya

Data Supply Chain Architect / Analytics Engineer · Product.ai

Owns the structured data layer that the company’s content pipelines, story generation, and verification engine all drink from — the engineer who decides what data hundreds of thousands of pages are built on every night.

İstanbul-trained analytics engineer turned data-substrate architect. If the data is wrong, thousands of bad pages get generated; when it’s right, every team downstream ships without re-asking the same question. Turns unstructured signals into reconciled, machine-readable pipelines — and ships the spec, the code, and the honest label in the same session.

LinkedIn → 3 Prompting Techniques That Actually Work →
Tenure
Joined Apr 2024 · 12+ year analytics arc
Based
Los Angeles, CA · via İstanbul
Owns
Ingestion · modeling · reconciliation · instrumentation
Prior
Yummly (Whirlpool) · Merkle · Hero Digital · N11 · İTÜ
30%
BigQuery cost reduction
$400K
Revenue opportunities surfaced
80%
Drop in ad-hoc analytics requests
12+yr
Analytics engineering arc
About

Data Supply Chain Architect / Analytics Engineer — the engineer who turns unstructured merchant signals into the reconciled, machine-readable substrate every other team builds on.

30%
BQ cost reduction
$400K
opportunities surfaced
80%
fewer ad-hoc requests
12+yr
analytics arc
The range

01 · The substrate, owned end-to-end

Data Ingestion

Federated reads across the commerce data warehouse, web-analytics telemetry, and search-console automation. The signals come from everywhere; she normalizes them into one place.

02 · The substrate, owned end-to-end

Modeling

A transformation layer built on modern data-modeling tooling — canonical tables, density scoring, self-serve columns reviewable without writing SQL.

03 · The substrate, owned end-to-end

Reconciliation

Cross-source bridges between systems that never agreed — manual trackers, warehouse row counts, adapter folders. Three sources of truth, one map.

04 · The substrate, owned end-to-end

Instrumentation

Fetchability canaries at startup, stratified sampling, history logs, multi-channel failure surfaces. Detection horizons compressed from weeks to seconds.

05 · Sharpened at Product.ai

Framework Authoring

Four load-bearing specs in 30 days. Specs ship with their self-critique cycle in the body, not the appendix — so future readers see why the thresholds are where they are.

06 · Sharpened at Product.ai

Production Verification

End-to-end testing behind authentication against live production — a reusable pattern for verifying any internal dashboard before it ships.

Marquee stops

From İstanbul engineering school to the data substrate behind the verification engine.

Apr 2024 → Now · Product.ai
Product.ai (formerly Demand.io)
Data Supply Chain Architect / Analytics Engineer
Architects and owns the analytics stack end-to-end on a $100M platform: BigQuery warehouse design, dbt layering, Prefect orchestration, GA4 taxonomy, and the canonical semantic layer 50+ stakeholders across 6 functions query directly. Operates inside Product.ai’s production AI operating system, building data products consumed by AI agents. Cut BigQuery cost 30%, surfaced ~$400K in revenue opportunities, and reduced ad-hoc analytics requests 80% via a conversational analytics layer (Claude over BigQuery via MCP). If the data is wrong, thousands of bad pages get generated.
Owns the data substrate
Before Product.ai
Jan 2022 → Mar 2024 · Remote
Yummly (Whirlpool Corp)
Business & Data Analytics Manager
Owned the analytics warehouse for a consumer subscription platform serving millions monthly. Canonical dbt models drove a 10% subscription revenue lift and cut ad-hoc reporting 30%. Built the experiment-readout framework that standardized A/B interpretation across Growth, Product, and Customer Success.
Warehouse ownership at scale
Sep 2018 → Dec 2021 · Remote
Merkle
Analytics Engineer
Led end-to-end analytics architecture for 5+ Fortune-500-tier healthcare and financial-services clients at once. Replaced one-off SQL and weekly reconciliation with governed Looker models. Built testing and governance in HIPAA environments.
Enterprise reconciliation
2017 → 2018 · San Francisco
Hero Digital
Digital Analytics Consultant
UTM governance, channel attribution, and cross-channel QA for enterprise clients across paid, organic, social, and direct.
Attribution & governance
2012 → 2017 · İstanbul
N11.com / Kariyer.net
First Data Hire
Founding analytics member at a major Turkish e-commerce marketplace. Architected the warehouse, modeling layer, and metric standards from zero through high-growth scale. Define the standard, do not wait for one.
Zero-to-one platform ownership
2007 → 2013 · İstanbul
İstanbul Technical University
M.S. Energy Sciences & Technology · B.S. Management Engineering
Six years at one of Türkiye’s flagship technical universities: undergraduate in management engineering, master’s in energy sciences and technology. The systems-engineering reflex — specify the substrate first, build the application against it second — traces back to here.
İTÜ · 2007–2013
The through-line

One job across every chapter: build the substrate other people can build on.

The work isn’t writing SQL — it’s removing the need for anyone to write SQL ad-hoc again. The North Star is the move from manual queries to autonomous data infrastructure: pipelines that refresh themselves so the teams downstream never have to re-ask the same question.

It carries a discipline most data work skips: if a version-one has a scope limit, label it — visibly, inside the interface, not buried in a footnote. Trust is earned by saying plainly what doesn’t work yet, not by claiming everything is supported.

And it treats AI the way it treats data: as a subsystem to govern, not an oracle to defer to. Current models act exactly like rushed humans — they skim, guess, and take shortcuts unless told otherwise. The same rigor that goes into a data model goes into every prompt.

Turn unstructured signals into the substrate other people can build on without re-asking the same question. Cansu, on the through-line of the work
Operating code

Six principles that show up across her shipping, her frameworks, and her published writing.

01 Method

Velocity comes from a substrate, not a query.

The work isn’t writing SQL — it’s removing the need for anyone to write SQL ad-hoc again. Manual queries become self-refreshing pipelines.

“By end of quarter, the generators run on self-refreshing pipelines with zero ad-hoc SQL from me.”

02 Architecture

One canonical source per thing. Always.

One canonical table standard, one runtime lookup pattern. Every reader uses one source; every writer goes through one path. No second source of truth.

“One canonical source per first-class entity — remember it in every session.”

03 Honesty

If version one has a scope limit, label it.

Scope limits ship inside the interface, not in a footnote. A methodology change gets pre-communicated to the people relying on it before they rely on it.

“Don’t frame it as everything is supported. Frame it as what actually is.”

04 Instrumentation

Measure what succeeded — not what failed silently.

Aggregate metrics quietly drop what was attempted-and-failed. A fetchability canary caught what weeks of dashboard reads missed — then got written into the framework so every future launch inherits the rule.

“What is the gate actually trying to catch, and what filters does that imply?”

05 Rigor

Critique your own output before you ship it.

Every framework spec carries its self-critique cycle inside the document. Future readers see why thresholds are where they are, not just what they are.

“Be rigorous. Critique your output and make it better. This part is incredibly crucial.”

06 Prompting

AI agents skim and guess unless you tell them otherwise.

Separate prompt-design from execution. Force self-verification through explicit checkpoints. Demand extreme explicitness — the same rigor she brings to data models.

“Current models act exactly like rushed humans: they skim, guess, and take shortcuts unless explicitly told otherwise.”

Scope

What she owns — and what she explicitly doesn’t.

Anti-goals are part of the role description, not a footnote. Cansu’s scope is the data layer — supplied to every other domain, executed in none of them.

Ingestion pipelines

Owns

Every signal, normalized into one place.

Federated warehouse reads, web-analytics streams, and search-console automation, joined into a single reconciled layer.

Modeling & reconciliation

Owns

One canonical source per thing.

A modern transformation layer, a canonical table standard, and cross-source bridges that reconcile systems which never agreed.

Framework authoring

Owns

The specs the queries get written against.

Content-density scoring, rollout-observation and analysis frameworks, reporting-QA — four load-bearing specs, each with its self-critique cycle in the body.

Audit instruments & verification

Owns

Failure surfaces, not silent gaps.

Fetchability canaries, stratified samples, history logs, multi-channel failure surfaces, and behind-authentication production testing.

Dashboards for humans

Anti-goal

Builds the data layer; humans build the surfaces.

She supplies the modeled, reconciled data — the visual surfaces on top of it are owned elsewhere.

Product, ops & editorial execution

Anti-goal

Supplies the substrate; doesn’t run the domain.

The verification engine, the operating model, and editorial content all drink from her data layer — she builds the layer, not the domains on top of it.

Published thinking

What sits on the open web.

LinkedIn · Article 2025

3 Prompting Techniques That Actually Work

Separate prompt-design from execution. Force self-verification via explicit checkpoints. Demand extreme explicitness. “Read 100% of this file line by line” — not “Review this file.”
LinkedIn · Post 2024

Hammer / Nail — orchestration vs. basics

Logging gaps, schema fixes, upstream design flaws — often easier to solve with basic engineering conversations than with intricate workflows. “Just because you have a hammer, doesn’t mean everything is a nail.”
What sets her apart

Six combinations rare individually. Unusual to find in one engineer.

Data substrate + framework author, simultaneously

Most analytics engineers ship queries. Cansu ships the specs the queries get written against — four load-bearing frameworks in 30 days, each consumed by another team’s downstream work.

The cross-source reconciliation reflex

Manual trackers, adapter folders, warehouse row counts — three sources of truth that never agreed, until she built the framework and the lookup that bridge them. One canonical source, made physical.

Stewardship threaded through every session

The gotchas memorialized, the bridge-keys saved, the chart annotations added. Cleanup isn’t a side activity — it’s woven into every shipping session.

The honest-labeling discipline

Scope limits live inside the interface, not in a footnote. A version-one ships with its limits visible. Trust earned by saying what doesn’t work.

AI as a subsystem to govern

Multi-agent orchestration at production scale — AI treated as a subsystem to govern, not an oracle to defer to. The published-thinking principle made code.

Istanbul engineering rigor at AI-commerce speed

Six years of engineering school, seven years in analytics, and dashboard fixes that collapse two-week timelines into two days. The depth and the velocity arrive together.

Career history and published writing on this page are publicly verifiable — LinkedIn, certifications, and university record. See the record →
The through-line

Turn unstructured signals into the substrate other people can build on without re-asking the same question — from the user-insights desk through the framework-author tier, shipping the spec, the code, and the honest label in the same session.

Product.ai builds with operators like Cansu — engineers who own a layer end-to-end and ship it with the honest label attached. See open roles →

Get in touch with Cansu See the team