
Sign up to save your podcasts
Or


When teams started shipping software continuously, security reviews couldn’t keep up. Developers needed feedback they could act on while the code was still fresh. FinOps faces a similar problem, and AI coding agents are making the development cycle even faster.
I sat down with Guy Podjarny, founder of Snyk and founder and CEO of Tessl, to talk about what it took to get developers to adopt security tools and what FinOps can learn from that experience. We also get into how teams can build security and cost checks into AI development workflows and improve those checks based on what happens in practice.
Key takeaways
Give developers feedback on the change they’re making. A pull request is a good place to flag an issue introduced by that change. An application-wide backlog can wait for a separate conversation.
Make the fix easy. Guy shares how Snyk used automated fix pull requests for low-risk dependency upgrades. For FinOps, the same idea starts with weighing the potential savings against the work and risk of making the change.
Treat FinOps as a service to engineering. Agree on cost expectations, then give developers the context and tools to meet them in their existing workflow.
Build feedback into AI development. Give agents clear requirements and access to the right checks, then use review feedback and evaluations to improve how they work as development speeds up.
Chapters
00:00 — Meet Guy Podjarny
00:50 — Why continuous delivery pushed security into development
06:17 — What drove Snyk’s adoption
08:37 — How security tools fit into developer workflows
11:52 — How security teams changed their relationship with engineering
16:24 — Reducing alert noise and making fixes easier
20:00 — Tessl’s approach to skills, feedback loops, and software factories
23:57 — When AI coding moves the bottleneck to code review
26:32 — How faster code generation changes security and cost management
31:12 — The role of security in AI development
33:47 — Keeping feedback loops as fast as development
About the guest
Guy Podjarny founded Snyk and is the founder and CEO of Tessl. He and I previously co-founded Blaze, which was acquired by Akamai. About Guy and Tessl
About the show
Shift Left FinOps is hosted by Mike Weider and explores how engineering teams bring cloud and AI cost decisions into the development process. Sponsored by Frugal.
An AI application that works can still be too expensive to run. So how do you decide what’s worth the investment?
Dieter Matzion returns to Shift Left FinOps to talk with Mike Weider about the economics of AI applications. Drawing on his work supporting FinOps at Roku, Dieter shares how a vendor-management tool went from $4 per question to 10 cents by replacing repeated document scans with cached summaries.
We explore how to measure cost per task, choose models, set spending limits, and ask business owners what an AI feature is actually worth. We also discuss how experimentation connects application costs to customer engagement and business outcomes.
Key takeaways
Measure cost per task. Token prices alone don’t tell you what answering a question or completing a workflow costs.
Cache and reuse work. Dieter’s vendor-management example shows how a change in application design can substantially reduce costs.
Match models to the task. Use evaluations to determine when a cheaper model delivers sufficient quality and when escalation is needed.
Give business owners the value decision. Engineering and FinOps provide cost visibility; business owners help define acceptable economics.
Account for volume before launch. Model calls, agent depth, and execution frequency all affect the eventual bill.
Test business impact. Evaluate whether a feature improves engagement, retention, or conversion alongside what it costs.
Chapters
00:00 — Tokenomics and the business-value question
00:32 — Welcome to Shift Left FinOps
00:57 — Dieter Matzion returns
01:27 — From $4 to 10 cents per AI question
05:33 — What tokenomics adds to the FinOps conversation
09:16 — The five layers of the AI cost stack
12:05 — Caching, model routing, and quality checks
15:20 — Setting spending limits and assessing business value
21:22 — AI overuse, API dependencies, and “shadow intellectual property”
24:39 — Estimating AI costs before launch
27:58 — Model choice, growing usage, and developer freedom
32:42 — Balancing model quality and cost
34:17 — Evaluations and testing frameworks
36:15 — Cost awareness before deployment and production monitoring
38:37 — Using A/B tests to connect cost with business outcomes
44:51 — Two tokenomics practices to start with
46:08 — Where to follow Dieter
About the guest
Dieter Matzion supports FinOps on Roku’s cloud and technology infrastructure team. He previously held senior engineering roles at Intuit, Netflix, and Google. Connect with Dieter on LinkedIn.
About the show
Shift Left FinOps explores how engineering leaders move cost optimization upstream into the development process. Hosted by Mike Weider and sponsored by Frugal, which helps engineering teams automate cloud and AI cost visibility and optimization before and after they ship.
Cost models help engineering teams understand the financial impact of a product or architecture before it reaches production. But what makes a model useful, how accurate does it need to be, and what happens when real usage differs from the assumptions?
In this episode, Mike speaks with Mariano Szklanny, Principal Operations Manager at Okta, about how cost modelling connects engineering decisions to pricing, margins, and business risk.
Mariano shares a real example where a cost model exposed duplicated fixed infrastructure costs early enough for the team to change the architecture. He also explains how Okta decides which projects require modelling, how models are built and validated across teams, and how production data is used to compare assumptions with reality.
What You’ll Learn
How a cost model changed the architecture of an Okta feature
The difference between a cost model, budget, forecast, and cloud calculator
Why models should begin with unit cost before forecasting total spend
How to identify the fixed and variable costs that matter most
How Okta decides which products and features require cost modelling
Why cost models are also communication tools for Engineering, Product, Finance, and Pricing
How AI reduced the time required to build a model from one or two months to roughly one or two weeks
Why Mariano replaced spreadsheets with interactive applications
How instrumentation allows teams to compare assumptions with production usage
Why cost modelling should enable business decisions rather than simply block spending
Episode Chapters
00:00 Why engineering teams need cost models
00:49 A cost model that changed an Okta architecture
03:06 What Okta does and Mariano’s path into FinOps
06:05 What is a cost model?
09:10 Cost models versus budgets, forecasts, and cloud calculators
11:38 How Mariano builds a cost model
13:28 Unit costs and adoption scenarios
14:19 How AI has changed the modelling process
16:35 How long does a cost model take to build?
18:05 Deciding which projects need modelling
21:10 How Product, Engineering, Pricing, and Finance work together
24:11 Instrumenting products and validating models after deployment
27:17 Balancing cost with growth, security, and reliability
30:16 Advice for teams getting started
About Mariano Szklanny
Mariano Szklanny is a Principal Operations Manager at Okta, where he combines a software engineering background, a Computer Systems Analysis degree, and nearly a decade at software consultancy Southworks with financial skills.
He builds product cost models that surface financial risks and opportunities early and help protect margin. Mariano joined Auth0 as Engineering Operations Manager in late 2015, reporting to the CTO and co-founder, and has been at Okta since its acquisition of Auth0 in 2021.
Cloud cost management found him at Auth0 when he was asked to investigate a runaway AWS bill. The results led Finance to bring him onto the team and expand the work from cost optimization into profitability. Today, he works across Product, Engineering, Pricing and Packaging, and FP&A, connecting engineering and product decisions to what they cost to operate at scale.
Mariano Szklanny on LinkedIn
This episode is sponsored by Frugal AI.
Shift Left: Lessons Learned from 30 Years of Building
With Dave North, Principal Engineer at Rewind
How do you make cost part of engineering without slowing engineers down?
Mike Weider talks with Dave North about how Rewind brings cost visibility and ownership into the software development lifecycle. Dave shares practical lessons on modeling costs before building, spotting inefficiencies in production, and getting engineers engaged through visibility and cost champions.
They also explore how AI is changing the process, from building interactive cost simulators to automating code reviews, and why small changes to prompts, logs, and database records can have a meaningful impact at scale.
About the guest
Dave North is a Principal Engineer at Rewind, a SaaS backup company serving more than 25,000 companies worldwide. He brings over 30 years of engineering experience across large corporations and startups. At Rewind, he has built and led CloudOps, platform, security, and IT teams. Today, he focuses on complex problems that span engineering and the wider business.
Key takeaways
Chapters
Links
Dieter Matzion, a cloud and technology infrastructure leader at Roku with prior engineering leadership roles at Intuit, Netflix, and Google, joins Mike Weider to talk about how FinOps teams can shift cost awareness earlier in the engineering lifecycle. T
he conversation covers why engineers often misjudge efficiency, how to create practical cost KPIs, and why AI is pushing the industry toward tokenomics as a separate discipline.
We discuss how to make cost another engineering dimension alongside latency and performance, why leadership buy-in matters, and how teams can keep cloud spend aligned with business growth without slowing down innovation.
Key topics
Timestamps
00:00 - Dieter Matzion’s background in cloud infrastructure and FinOps
00:26 - Where shift-left FinOps stands in the community today
03:35 - Why engineers are often poor judges of efficiency
05:38 - The “good enough” mindset and why it fails for cloud commitments
07:25 - Incentives, downside risk, and why savings are rarely rewarded
09:50 - Making cost a normal engineering optimization dimension
10:47 - Why cloud cost literacy matters across the application lifecycle
12:32 - Basic KPIs teams can start with, from cloud cost per revenue onward
13:23 - Cost per vCPU hour and cost per gigabyte stored
15:09 - Why unit economics must be measured in production, not just planned
16:02 - A real example of AI tuning reducing monthly cost dramatically
16:23 - How a gossip protocol created a major cost explosion
16:50 - Roku’s cost per streaming hour benchmark
19:52 - Using proof of concept work to estimate material cost impact
20:53 - How cost checks happen at rollout level, not just per individual code change
21:53 - Alerts, escalation, and human-in-the-loop response
22:57 - Pure cost anomalies versus KPI-driven cost regressions
23:34 - Why tokenomics exists alongside FinOps
24:04 - FinOps versus tokenomics across the AI stack
25:46 - The five layers of the AI stack and how they change optimization
26:11 - New tokenomics KPIs like cache hit ratio and routing quality failure rate
27:35 - Why AI requires a distinct discipline beyond traditional FinOps
28:40 - How AI coding tools and AI application spend change the shift-left problem
29:19 - Current enterprise limits on AI usage and who drives most spend
30:54 - How agentic workflows could eventually include a cost-aware FinOps agent
31:49 - Best advice for teams in the crawl phase
32:10 - Start small, build successes, and get leadership buy-in
33:28 - Why Roku created cost per streaming hour to keep spend from outpacing revenue
From the publisher's feed