ShipTalk · 2025-12-01 · 45 min
Key moments - from our scoring
Substance score
68 / 100
Five dimensions, 20 points each
Marina Petzel brings eight years of experience shipping ML systems across domains - from oil production to healthcare research at UCLA to 2D CAD design at Autodesk. Her current work focuses on Smart Blocks, a computer vision feature that detects and recommends geometries in AutoCAD drawings, automating the manual copy-paste workflows that have existed for decades. Rather than chasing state-of-the-art models, Petzel emphasizes the unglamorous reality: roughly 70% of development time goes to data collection, cleaning, and infrastructure - not model training. She walks through Autodesk's annual release cycle, where teams have one year to move from ideation through prototyping, validation, model training, and shipping. Petzel argues that ideation is deliberately underestimated in the AI era, where teams often force ML into problems that don't need it. Her approach to AI productivity centers on extending human skill sets rather than replacing them, while stressing the importance of prompt engineering, red teaming, and responsible guardrails when deploying generated code and AI-assisted features. For B2B operators shipping ML products, her insights on user shadowing, fast prototyping with simple models, and balancing development speed with safety testing offer a grounded counterpoint to AI hype.
Approximately 70% of development time is spent on data preparation, cleaning, collection, building pipelines, and infrastructure; only a smaller portion goes to model training and fine-tuning.
SmartBlocks is a computer vision feature that detects and recommends geometries (doors, furniture, symbols) in 2D CAD drawings, automating the manual copy-paste workflows users have relied on for decades and helping them find and insert repetitive elements faster.
Autodesk releases new features every March; the team has one year to move from ideation through extensive research, prototyping, user validation, model development, evaluation, and production deployment.
User shadowing and research (including YouTube videos of user workflows) help engineers understand the actual repetitive and painful parts of a user's job, which is especially critical if the engineer lacks domain expertise like architectural knowledge for CAD.
The team prioritizes data quality (garbage in, garbage out), requires rigorous testing for model predictability and unexpected behavior, and emphasizes that testers and QA professionals are essential to catching violations and ensuring models work as expected before release.
Our reviewer’s read on each dimension, with quotes from the episode.
The episode delivers solid practitioner insight on ML development workflows, particularly around the ideation-to-production cycle at Autodesk. Marina provides concrete process guidance (70% of time on data prep, prototype validation before scaling) and challenges common misconceptions (don't use state-of-the-art models at prototype stage). However, the conversation drifts into general AI adoption platitudes and mentoring advice that are less novel for experienced operators. The content is substantive but not densely packed with non-obvious claims.
about 70 percent of your time you will spend preparing your data, cleaning your data, collecting it, building pipelines, and also spending a significant amount of time on infrastructure
the prototype stage your main goal is just to prove that your problem and your solution are valid and make sense
Marina offers a practitioner's perspective on real ML deployment constraints that diverges somewhat from hype-cycle narratives (emphasizing ideation, data quality, staged validation). Her point about shadowing users and understanding problem spaces before solution-building is sound but not particularly novel. The discussion of responsible AI adoption, multi-agent systems, and data privacy are contemporary but largely echo existing best practices rather than introducing fresh frameworks or counterintuitive arguments.
especially now in the age of AI, where everyone tries to put AI in every single thing
the last thing that really surprised me was a laundry machine with "AI features," and I was like, I do not know what that is exactly
Marina is a credible practitioner with 8 years of ML experience across multiple domains (oil production, healthcare, CAD), currently shipping production ML features at scale at Autodesk. She leads an internal AI adoption working group and actively mentors engineers. She is not a pure research or thought-leader type - she has shipped features yearly and understands production constraints. However, she is not a founder or C-level operator defining strategy at scale, which limits the score slightly.
I work as a Senior Machine Learning Engineer and also AI Productivity Lead at Autodesk. I joined Autodesk about four years ago
For the last three years, we shipped three different features
Marina provides concrete details about AutoCAD's 42-year legacy, the yearly March release cycle, the Smart Blocks feature for detecting and recommending geometries, and the breakdown of work (70% data prep, model development, shipping). She mentions specific tools (Cursor, GitHub Copilot, Claude, Replit, Lovable) and naming Autodesk as her employer. However, she rarely cites metrics, dollar figures, user adoption numbers, or quantifiable business outcomes. The medical records RAG example is personal, not a deployed product case study with measured results.
One of the features that I've worked on for the last three years is Smart Blocks, and this feature is about detecting and recommending geometries
AutoCAD. It's one of the oldest products in Autodesk, which is, I believe, about 42, maybe 43 years old
Dewan asks thoughtful, substantive follow-up questions (e.g., on safety, responsible adoption, testing imbalance) and occasionally challenges assumptions (e.g., questioning whether testing can keep pace with AI-generated code). However, he rarely pushes back on Marina's claims or probes disagreement; most exchanges are confirmatory. Marina's answers are articulate but rarely interrogated deeply. The conversation feels collaborative and friendly rather than adversarial or deeply exploratory. Some softer moments (e.g., general advice on mentoring) lack sharp follow-ups.
How do you ensure that the model is safe to use, especially for customer-facing features?
Do you see an imbalance in that mindset, where you have a ton of code being written, some by human users, some by agents, but then who is testing all this code?
Computed from the transcript - who did the talking, and the words that came up most.
In this episode of Ship Talk , host Dewan Ahmed (Principal Developer Advocate, Harness) sits down with Marina Petzel , Senior ML Engineer and AI Productivity Lead at Autodesk, to unpack what it actually looks like to ship AI into long-lived, production software. Marina shares her journey from classic predictive analytics to computer vision and LLMs, and how her team brings intelligent features like Smart Blocks in AutoCAD from ideation to prototype to production on a tight yearly release cadence. She breaks down the often ignored parts of ML work: data pipelines, infrastructure, UX fit, and safety testing, and explains why “garbage in, garbage out” is a law of nature, not just a cliché. They also dive into AI for developer productivity , multi-agent workflows, red-teaming and prompt safety, and how every engineer, regardless of seniority, can start using AI as a genuine skill multiplier. Marina closes with concrete advice for aspiring ML engineers: build more, read less, follow your curiosity, and share your work publicly so your visibility compounds over time. Follow Marina's work:
Transcribed and scored by The B2B Podcast Index.
WEBVTT 1 - > Dewan Ahmed: Welcome to another episode of the ShipTalk Podcast, where we talk about the ins and outs, ups and downs of software delivery. Today's episode will be a treat. 2 - > Dewan Ahmed: If you're curious about how AI is actually used in real products, not just in blog posts and hype cycles, our theme today is Shipping Practical AI, LLMs, Productivity, and Real-World ML. 3 - > Dewan Ahmed: To explore that, we're joined by someone who lives and breathes this daily, Marina Petzel, Senior ML Engineer and AI Productivity Lead at Autodesk.
Marina works at the intersection of machine learning and engineering enablement, 4 - > Dewan Ahmed: and organizational transformation. She builds customer-facing ML features, drives AI adoption inside teams, and shares her knowledge widely online. Marina, welcome to Ship Talk. 5 - > Marina Petzel: Thank you for having me, it's very exciting to be here.
6 - > Dewan Ahmed: Likewise, let's start with a quick intro for our listeners who may not have seen your work yet. Who are you, and what do you spend most of your time on these days? 7 - > Marina Petzel: Sure, that's a big question. My name is Marina Petzel, I'm based in the Bay Area, California.
8 - > Marina Petzel: And as you said, I work as a Senior Machine Learning Engineer and also AI Productivity Lead at Autodesk. I joined Autodesk about four years ago. 9 - > Marina Petzel: Before that, I worked for different companies as a data scientist. I also worked as an AI researcher for UCLA, where I did research in the healthcare industry.
10 - > Marina Petzel: Over the last eight years, I would say, I've worked across different domains in data science, starting with the oil production industry, and now I'm working in 2D CAD drawings. 11 - > Marina Petzel: I started with classic predictive analytics models, back when transformers were not a thing. Now I have shifted more toward computer vision and also LLMs, as many of us have today. 12 - > Marina Petzel: So yeah, I think that is primarily it.
13 - > Dewan Ahmed: And what sparked your interest first? A lot of people are trying to get into this field, and sometimes they are not even from, let's say, a computer science background. 14 - > Dewan Ahmed: They are wondering how other people moved into this field. Of course, you are from a technical background, technical education, but was there a point in your career where you thought, 15 - > Dewan Ahmed: “Oh no, this is what really interests me, and I want to go this route”?
16 - > Marina Petzel: Yeah, actually, when I graduated from my bachelor's, in the last year, 17 - > Marina Petzel: my bachelor was in computer science, and I started to think what I wanted to do next. I realized pretty early that traditional software engineering didn't quite excite me. 18 - > Marina Petzel: And at that time, like I already mentioned, about eight years ago, I started to think what it could be. Data science and AI were not what they are today, so I discovered them more from a predictive analytics angle and started to do a lot of research.
19 - > Marina Petzel: My capstone project was related to that. I tried to build a stock-trading robot using reinforcement learning. 20 - > Marina Petzel: That was pretty exciting for me. There were not a lot of professionals around me at that time, so it was pretty new in many ways.
21 - > Marina Petzel: And that was when I realized 22 - > Marina Petzel: that everything new excites me. So I started to dive deeper into it. Then came the time where I went to grad school, a master's in data science, and I took a computer vision class. 23 - > Marina Petzel: That honestly completely changed my trajectory.
I loved the combination of math, algorithms, and visual intuition behind it. 24 - > Marina Petzel: Since then, I joined Autodesk and started to work on computer vision systems. 25 - > Marina Petzel: We build a lot of different features that are customer-facing for CAD drawings, and it's really exciting to see that these features are very useful for the users. So that is my trajectory.
26 - > Marina Petzel: For everyone who is starting today from scratch, I would say that, for me, this might be an even easier route, because right now there is definitely a big demand for AI. 27 - > Marina Petzel: But there are also some areas where everyone is new. It is really good to start somewhere where no one knows what the right answer is, and it is okay to experiment, to make mistakes, and to try different things. 28 - > Dewan Ahmed: I love that, that idea that everyone has something to learn.
This is not an established field yet like math or physics, which have hundreds of years of research behind them. AI is moving pretty fast. What excites you about this current momentum? 29 - > Marina Petzel: Yeah, 30 - > Marina Petzel: I mean, as I already said, I'm really excited that it's new for everyone, and we can imagine that we are all in some sort of sandbox right now, and everyone can build a prototype.
Even 31 - > Marina Petzel: UX designers or people who were completely outside of the technical and coding world can build a very cool prototype within two days and show it to their peers. 32 - > Marina Petzel: That truly excites me, that AI is becoming accessible for everyone, 33 - > Marina Petzel: for every generation. I just came back from visiting my parents, and they are using ChatGPT and DeepSeek all the time. I set up GitHub Copilot for my dad, and I told him what AI-assisted coding is, and that excites me, that we are all new to it and we all 34 - > Marina Petzel: have an opportunity to try it and see how we can get the best use out of it.
35 - > Dewan Ahmed: That part is undeniable. From my realtor telling me how he now gets all his work done with GPT tools to the doctor's office having a note that they use AI tools, 36 - > Dewan Ahmed: you can see the widespread adoption and the fast pace, even in industries that are usually slow movers. 37 - > Dewan Ahmed: You mentioned customer-facing features and customer-facing ML work, and that is something we typically do not hear that much about. We hear more about research and what is coming next.
Would you be able to share with our listeners some of the customer-facing ML work you have been doing? 38 - > Marina Petzel: Yeah. 39 - > Marina Petzel: Absolutely. As I mentioned, I'm working on CAD 2D drawings.
The product that I'm working on is called AutoCAD. It's one of the oldest products in Autodesk, which is, I believe, about 42, maybe 43 years old. 40 - > Marina Petzel: So there are a lot of different established tech processes behind this software, but we also started our machine learning team to help our users 41 - > Marina Petzel: make their workflows more intelligent. One of the features that I've worked on for the last three years is Smart Blocks, and this feature is about detecting and recommending geometries.
42 - > Marina Petzel: For those of you who are not familiar with what AutoCAD is used for, you 43 - > Marina Petzel: can think about floor plans. Some of us have definitely looked at real estate or seen different floor plans, where we have walls as a frame, but also different rooms, 44 - > Marina Petzel: and geometries inside of them, like doors, furniture, or different symbols. I am working on optimizing the manual workflow for our users, optimizing how they search and how they insert these repetitive geometries.
45 - > Marina Petzel: Before we started working on that, like 42 years ago or even 20 years ago, all the workflows were manual copy-paste. 46 - > Marina Petzel: People would scroll through a lot of different drawings in order to find some essential objects they needed. Now we put machine learning models, machine learning algorithms behind this, which help 47 - > Marina Petzel: our users find the geometries they need faster, and also identify geometries 48 - > Marina Petzel: that might be broken or that require the user to do some additional steps to make the best use of them.
49 - > Marina Petzel: That is what I'm working on, and for the last three years, we shipped three different features. 50 - > Marina Petzel: In AutoCAD, it's also very interesting that every March we have a new release, so every year we need to ship something. Every year we try to surprise our users with new features and make their workflows smarter, 51 - > Marina Petzel: not harder. 52 - > Dewan Ahmed: That is fantastic to hear.
For those of our listeners who are just entering the field of machine learning and data science, maybe they are familiar with, say, “How does a computer know it is a cat?” You show it a thousand cat pictures, and then it learns the features of a cat picture. 53 - > Dewan Ahmed: What is your process like for AutoCAD? You mentioned automating some of the manual tasks your users have been doing, so what is your process from ideation to prototype to production?
54 - > Marina Petzel: As I mentioned, we have a release date. We know it is going to happen every year in March, so we do not have a lot of time. We just have one team working on all of these features. 55 - > Marina Petzel: We have one year to come up with an idea, do some extensive research on it, prototype, validate it, 56 - > Marina Petzel: train the model, evaluate it, and ship it to production.
There is not a lot of time, which is why 57 - > Marina Petzel: we start preparing much earlier for what we are going to work on next year. It starts, obviously, with the ideation process. That process is always underestimated, in my opinion, 58 - > Marina Petzel: especially now in the age of AI, 59 - > Marina Petzel: where everyone tries to put AI in every single thing. The last thing that really surprised me was a laundry machine with “AI features,” and I was like, I do not know what that is exactly, and I am not sure I need that automated yet.
60 - > Marina Petzel: That is why it is really important to spend a sufficient amount of time on this ideation part and think about 61 - > Marina Petzel: not 62 - > Marina Petzel: what ML model we can put here, but what type of problem we want to remove for our user. 63 - > Marina Petzel: What is the most repetitive and painful part for them that we can optimize? 64 - > Marina Petzel: For that, I find it really helpful to shadow users if possible, or do some research on YouTube and watch different people talking about their experience.
Because I am not a designer, I am not an architect by background, so I do not have that type of 65 - > Marina Petzel: domain experience to understand the user directly. That is why it is very helpful, not just for PMs or UX designers 66 - > Marina Petzel: to explore the user's problem space, but for every engineer to walk through that and shadow this type of process. 67 - > Marina Petzel: Once we define the problem, the next step is obviously to start prototyping it and think how we can solve it.
As an ML team, we always try to make fast prototypes. Usually it takes 68 - > Marina Petzel: a few sprints 69 - > Marina Petzel: to build something pretty simple. We do not try to build sophisticated models on huge amounts of data at that point, and that is a mistake I see a lot of teams make. They try to 70 - > Marina Petzel: use state-of-the-art models, the latest and greatest, and that is usually a mistake, because at the prototype stage your main goal is just to prove that your problem and your solution are valid and make sense.
71 - > Marina Petzel: The goal of the prototype stage is, when you create it, 72 - > Marina Petzel: to understand if there is a signal here and whether users understand and actually want the behavior your model is producing. That is why, usually after the prototype phase, we try to do a 73 - > Marina Petzel: feedback loop with our users and understand whether it is something they are looking for or not. Maybe we should instead solve a different problem. 74 - > Marina Petzel: If we get all the green lights here, we move to one of the biggest parts of our annual life cycle, which is development.
75 - > Marina Petzel: And development, for some, 76 - > Marina Petzel: I know a lot of experienced data scientists definitely know this, but some people who are new to the field probably think development is just training some very fancy machine learning model, trying this model or fine-tuning that model. 77 - > Marina Petzel: I may disappoint some of you, but actually 78 - > Marina Petzel: about 70 percent of your time you will spend preparing your data, cleaning your data, collecting it, building pipelines, and also spending a significant amount of time on infrastructure.
79 - > Marina Petzel: That is what takes most of our time. There are definitely a lot of iterations back and forth, where you define metrics and try to optimize them, 80 - > Marina Petzel: and do some tweaking of your model. That is totally normal. Sometimes, when you start tweaking, you realize 81 - > Marina Petzel: it might not work the way you expected.
I would suggest that all teams be flexible and able to share these insights with stakeholders, so you can 82 - > Marina Petzel: make quick decisions and adjust your solution based on new insights you discover during the training phase. 83 - > Marina Petzel: The last phase, I think, is my favorite. Obviously, when all the dirty work is done, the model is ready, it performs the way you want, and you are preparing it for shipping to your customers. 84 - > Marina Petzel: That is kind of the annual cycle we are all working in.
85 - > Dewan Ahmed: That is fascinating to hear, and I think you answered a lot of the questions that I was going to ask next, such as what some of the lessons learned were. You mentioned the importance of good UX, that “measure twice, cut once” mentality, where you 86 - > Dewan Ahmed: decide first, understand it, and do not immediately jump into building. You talked about model - user mismatch. You also talked about the common misunderstanding that you are not going to be working on fancy models initially.
A lot of time is spent on cleaning and preparing the data. 87 - > Dewan Ahmed: I want to highlight one other important part: safety. How do you ensure that the model is safe to use, especially for customer-facing features? How do you ensure the safety part of those models?
88 - > Marina Petzel: Yeah, that is a really good question. As I said, one of the biggest chunks of the annual development cycle is preparing your data. 89 - > Marina Petzel: That is one of the first and most fundamental things that can affect your final model’s safety. We never sacrifice data quality.
The statement is simple and you all heard it: garbage in, garbage out. It is not a cliché in ML, 90 - > Marina Petzel: it is truly like a law of nature, because a lot of quality gains come from better data curation. 91 - > Marina Petzel: From the safety standpoint, obviously, we are all vulnerable now. All of us are using tools like Cursor, GitHub Copilot, Claude, 92 - > Marina Petzel: and a lot of code is starting to be generated using AI, and model safety is becoming even more important.
That is why, at the end, when I say my favorite part is when we are ready to ship the model, 93 - > Marina Petzel: it is my favorite part as a developer, but it might not be the favorite for testers, the people who are on the front line testing the models and making sure that the behavior 94 - > Marina Petzel: is not violating any rules and works as expected. Obviously, ML engineers and researchers all do this type of testing, 95 - > Marina Petzel: but once you are shipping a customer-facing feature, you need to make sure 96 - > Marina Petzel: your model is predictable to the greatest extent possible and there are no surprises along the way.
97 - > Dewan Ahmed: Yeah, I totally hear that, because at the end, it is similar to working in teams, where you are building something, someone is testing, someone is delivering, but for the end product, everyone is responsible for the quality. Everyone's work is equally important, from the person getting the data, even interviewing someone, to the person 98 - > Dewan Ahmed: who writes the release docs. 99 - > Marina Petzel: Right, right. 100 - > Dewan Ahmed: So, switching gears, part of your title is AI Productivity Lead.
What does AI for productivity mean to Marina? 101 - > Marina Petzel: That is a good question. AI for productivity, in my view, is 102 - > Marina Petzel: about the skills we all have. Someone knows how to play musical instruments, someone is good at coding.
We all develop these skills, and it is hard work. 103 - > Marina Petzel: But right now, since AI is getting more and more capable, for me, 104 - > Marina Petzel: it is about how you can extend your skill set by leveraging AI. 105 - > Marina Petzel: Some people are opening completely new angles of their personality by changing their day-to-day routines with AI. I see some of my friends actually using AI as their health coach or fitness coach.
For me, 106 - > Marina Petzel: it is very important to communicate that AI right now is not just an additional or nice bonus for some engineers or people who work in tech. 107 - > Marina Petzel: We have already crossed the line where AI can be used by everyone. 108 - > Marina Petzel: You might not have discovered yet how you can use it in your life, but it is definitely something that can 109 - > Marina Petzel: increase your productivity and even increase the quality of your life at some point.
110 - > Dewan Ahmed: Yeah, totally. One other thing I can think of is the amount of code that is being generated. That is probably like 10x or 100x, we do not know how many x, but 111 - > Dewan Ahmed: in the past, we had a team of engineers writing code. In the traditional setup, we had testers or QA engineers testing the code.
112 - > Dewan Ahmed: Do you see an imbalance in that mindset, where you have 113 - > Dewan Ahmed: a ton of code being written, some by human users, some by agents, but then who is testing all this code? 114 - > Marina Petzel: Yeah, one of the good use cases of AI is also writing these tests. But I am not the type of person who is completely bullish on AI and thinks that it is a magic wand that will help us 115 - > Marina Petzel: everywhere. Obviously, responsibility is still on us and how we use these great tools.
116 - > Marina Petzel: That is why it is very important not to just blindly accept all recommendations from AI, but to have a sense of what it does for you and what potential risks it can introduce to your systems. 117 - > Marina Petzel: I think we are all walking along some learning curve, some of us at the beginning, some of us already further along and adopting new advanced features pretty fast. 118 - > Marina Petzel: But it is getting clear that we need to be diligent about how we collaborate with AI.
119 - > Marina Petzel: Prompt engineering was hyped about one or two years ago, where prompt engineer salaries were just huge and everyone wanted to do that. 120 - > Marina Petzel: But that was for a reason. Prompt engineering is not just a straightforward thing. It is a truly powerful tool that can help you 121 - > Marina Petzel: avoid some situations you do not want to be in, like security issues.
You definitely need to address that in your prompts, in addition to other guardrails as well. 122 - > Marina Petzel: Testing is a very important part, and especially with the greatest tools also come the greatest risks. We see a lot of focus around red teaming, right, like how we can catch 123 - > Marina Petzel: prompt injections or prompt jailbreaks. We need to talk about this with everyone, and again, as I mentioned, we might all be at different levels of this learning 124 - > Marina Petzel: curve of adopting AI.
That is why I think it is a responsibility for those of us who have already adopted and are using AI in many ways in our lives. We need to share that with everyone and 125 - > Marina Petzel: explain how we can get the best use out of it without introducing vulnerabilities, obviously. 126 - > Dewan Ahmed: We definitely do not want to get into vulnerabilities, because with great power comes great responsibility, and with responsibility we can also think about adopting AI responsibly.
That is one of the things you do, helping engineering teams adopt AI responsibly. 127 - > Dewan Ahmed: How does responsible AI adoption look? I want to add some context to that, because a lot of engineering teams think, are we too small for that, maybe it is only for large engineering teams and large engineering orgs. 128 - > Dewan Ahmed: What is your understanding of responsible AI adoption, regardless of the size of the engineering team?
129 - > Marina Petzel: Everyone can start with something simple: data privacy, and thinking about what you expose to your AIs, and making sure that you put the right guardrails in place. 130 - > Marina Petzel: We take potential threats like prompt injections very seriously. We do some red teaming and define 131 - > Marina Petzel: clear prompting patterns. Again, I am leading the working group at Autodesk for AI for Productivity, 132 - > Marina Petzel: where we realized that our mission right now is to create a library of prompts.
We all 133 - > Marina Petzel: have different experience with AI, some of us are just starting, but 134 - > Marina Petzel: at the same time we are all in the same boat. We want to make sure that we are secure and that we do not expose the company or our customers’ data to risks. 135 - > Marina Petzel: That is why our group is working on creating this space where everyone is able to share their best practices, 136 - > Marina Petzel: for example prompts that are more specific to certain use cases, to the tool they are using.
We want to make sure that everyone who is just starting or does not really know how to frame a prompt can leverage that. 137 - > Marina Petzel: That is definitely something we can 138 - > Marina Petzel: help with. Also, we are embracing building internal MCP servers that check for security issues, and I think there are a lot of 139 - > Marina Petzel: lessons that you can find on Medium or different web sources, best practices on how you can make sure that before you push your code 140 - > Marina Petzel: it passes all security checks.
I think that is one of the key things all developers need to think about. 141 - > Marina Petzel: If you are not a developer, do not think that it does not apply to you. You also need to think about what you feed to your AI. When you talk to any GPT model, make sure that you do not expose any sensitive information and that you cannot be compromised with that later.
142 - > Dewan Ahmed: Totally, and I think these days we are all builders, regardless of whether we are software engineers or product managers. Everyone 143 - > Dewan Ahmed: has that tool and can try to build things. At the end of the day, the responsibility falls equally. 144 - > Marina Petzel: Absolutely.
Totally agree. 145 - > Dewan Ahmed: You have been very hands-on with multimodal models and chatbots. What are some of the interesting use cases you have seen here? 146 - > Marina Petzel: Again, I am primarily working with computer vision features, but definitely when, two and a half or three years ago, ChatGPT 147 - > Marina Petzel: really took off, it sparked a new interest in building things.
I played with different RAG models, multimodal RAG models, where we can 148 - > Marina Petzel: introduce visual capability as well as text, and try to understand the CAD space, 149 - > Marina Petzel: the drawings or designs, better. It is really good to see that it is already reasoning over these drawings and trying to associate notes and specifications. 150 - > Marina Petzel: Lately, what I played with, and I think it might be very interesting not only for our tech audience but also for non-technical people, is how to set up multi-agentic 151 - > Marina Petzel: workflows.
I played with a few systems to set up multi-agents, like Copilot Studio and also ChatGPT. 152 - > Marina Petzel: It is a very powerful approach, definitely not free of mistakes, and the hallucination level is still something that needs to be improved. 153 - > Marina Petzel: But you can organize your agents to specialize in different tasks. Some agents might be responsible for sending emails to customers, some 154 - > Marina Petzel: agents will be responsible for answering questions about the product, and that is really cool, because you can visualize the whole system and how it interacts.
155 - > Marina Petzel: I think many organizations - all organizations - will need to adopt this type of approach at different levels, because 156 - > Marina Petzel: at the end of the day, especially now, the speed of producing data and knowledge has increased, and now the problem is how I can figure out what I should focus on. 157 - > Marina Petzel: I also find these types of systems helpful in terms of searching. If something happened in the past, for example, if I am working on a product that is 42 years old, 158 - > Marina Petzel: and I am trying to find the problem space for me, I can look at past documents, what we discussed, and why we might not have pursued a specific route or solved a specific problem earlier.
159 - > Marina Petzel: Multi-agent systems are definitely the future. They might not be ideal yet, so I do not 160 - > Marina Petzel: want to raise expectations too high for our audience; they are definitely still prone to mistakes. But even with ChatGPT, remember what we used three years ago, when GPT answered “I only know everything 161 - > Marina Petzel: up to a certain year.” Right now, we do not have that as an issue in the same way.
162 - > Marina Petzel: The speed is fast, and it is very exciting. It might be overwhelming as well, but I think this is something we can already start using to change our lives and our work. 163 - > Dewan Ahmed: Yeah, I remember the early days with questions like “How many r’s are in strawberry,” and those kinds of prompts. Now I think many people are saying, “Let’s have ChatGPT build me a website,” even though 164 - > Dewan Ahmed: that person may have never opened an editor.
It is a very different, very fast pace in how people are even setting expectations. Legacy products, right, used to mature over years or decades, and now new products mature at the speed of light. 165 - > Marina Petzel: Yeah, totally. Again, I am very excited about the time that we are living in, and I think it is a really good time for everyone who is new to the industry to enter it now, because you can just spend some time prototyping using live coding tools 166 - > Marina Petzel: and get some knowledge.
Obviously, you should avoid the feeling that if you prototype one website that is not quite working but looks fancy, 167 - > Marina Petzel: it does not mean you already know everything about AI. But definitely, if you find a niche that fascinates you and try to dive deeper, 168 - > Marina Petzel: for example prompt engineering, and you read a few papers, experiment by yourself, and play with different prompting techniques, 169 - > Marina Petzel: you will already know more than, I would say, even one percent of the world.
This is a fascinating time where you can enter the field and become one of the top experts in a niche. 170 - > Dewan Ahmed: Yeah, and one way they can do that is by reading your technical content. Marina produces a lot of technical content, she is active on Instagram. What inspires you to share publicly?
171 - > Marina Petzel: Outside my day job, I like to share some of my AI insights. Again, I am leading an AI for Productivity group, 172 - > Marina Petzel: and there we explore different angles of using AI coding tools like Cursor, GitHub Copilot, Claude. 173 - > Marina Petzel: We also leverage different AI techniques to make developers more effective. What I realized after working and talking with many developers is that 174 - > Marina Petzel: I had to answer the same questions over and over again and explain the same fundamentals.
It is not bad, we are all on this learning curve and may be at different stages. But at the same time, I realized that if many people 175 - > Marina Petzel: do not have answers to these types of questions, maybe it is worth sharing my thoughts with a broader audience. So I focus a lot on how to do better prompting, how to prepare your data, how to 176 - > Marina Petzel: make the best use of AI coding tools right now, and what techniques you can use. 177 - > Dewan Ahmed: I love that, and we will be sure to link her Instagram account, websites, and any other public links in the description.
You also mentor a lot of junior ML engineers. What would be the three common pieces of advice you see coming up in your discussions? 178 - > Marina Petzel: I constantly hear that young professionals do not have the confidence to enter the field, because they think they need to have years of experience working in the AI field. Again, I want to emphasize that this is not true anymore.
179 - > Marina Petzel: Just try to start building. Build more, read less. Honestly, right now there is such huge data overload, it is hard to focus and get the best out of it. 180 - > Marina Petzel: I think right now you can build something pretty fast, so you will not get bored or feel 181 - > Marina Petzel: disoriented, because AI coding tools are really helping you build prototypes quickly.
On that excitement, you can build something that truly fascinates you. I always recommend starting with a project that personally interests you 182 - > Marina Petzel: and start building it. For example, when I started working with RAG and exploring it by myself, I had a personal problem I wanted to solve. I have lived in many different countries and accumulated a lot of medical records in different languages, 183 - > Marina Petzel: with different measurement units, etc.
I wanted to build a system that was slightly more complicated than just RAG by itself. It also involved some OCR, but I started to build it to solve my problem. 184 - > Marina Petzel: That also keeps you going, because you have a personal interest in resolving it, and I think that keeps you excited for a while. You should not wait until you feel, “Okay, I am ready to build something because I read these 185 - > Marina Petzel: five books about AI.
” No, that is not true anymore. That is why I recommend my mentees 186 - > Marina Petzel: embrace tools like Cursor, GitHub Copilot, Claude, Lovable if you are less technical, Replit, and so on. Use any tools you want and try to build as much as you can. Also, 187 - > Marina Petzel: what is very important, especially in the age of social media, AI, and content creation, 188 - > Marina Petzel: is to feel free to share your learnings publicly, because visibility actually compounds over time.
You will never know who will see your post or whether you will find partners to ideate on new ideas with. 189 - > Marina Petzel: So feel free to share your learnings, your thoughts, your lessons learned, even though it might not get a lot of likes or anything. You are doing it first of all for yourself. Try to document your learnings as blogs or GitHub repos.
That is a perfect 190 - > Marina Petzel: way to learn and stay motivated. That is one of the most important parts. 191 - > Dewan Ahmed: Could not agree more, totally. You have also been recently nominated as an engineering influencer.
What does that feel like? What does this label mean to you? 192 - > Marina Petzel: In terms of being seen as an engineering influencer, I do not try to think of it as being loud. I think of influence 193 - > Marina Petzel: as enabling others, and I think 194 - > Marina Petzel: we should talk about things and share them with others.
That is how I feel empowered, seeing that more and more people start using these tools. 195 - > Marina Petzel: They find them very helpful, and it is especially motivating when someone 196 - > Marina Petzel: shares personal stories about how you impacted their life or were an inspiration for building something. I am mentoring people from multiple parts of the world, and it is very exciting to see their success stories. A lot of people come from completely different fields, they never used AI or ML, and now they are already 197 - > Marina Petzel: shipping products within their companies.
That is very exciting and makes me feel happy. I think AI is not going to replace us, but it will make us more capable of doing things. 198 - > Dewan Ahmed: Totally. We are almost at the end of our podcast.
What would be one AI practice every engineer could try this week? 199 - > Marina Petzel: Go ahead and build something. I would say, if you are new to the field, definitely go and find something simple that you want to do. A quick example: I had a one-on-one with a non-technical colleague recently, 200 - > Marina Petzel: and we prototyped an Airbnb-like website, a booking-style website, within five minutes.
201 - > Marina Petzel: That was during the call. He was completely unaware of what live coding was, and it was possible to do in five minutes. So if you are new to the field, 202 - > Marina Petzel: please feel free to download a tool. There are a bunch of them.
If you are non-technical, use Replit, use Lovable, just go ahead and prototype and try to feel it. I think that will excite you and push you to move forward. 203 - > Marina Petzel: If you are already an experienced engineer or developer, try to explore something new for you. If you have never set up a multi-agent system, go ahead and figure out how to do that.
Again, 204 - > Marina Petzel: technology and products are developing so fast, and you constantly need to learn every week what is new. If we shift our mindset to following our curiosity, 205 - > Marina Petzel: I think that will definitely bring big benefits to all of us and to the world in general. 206 - > Dewan Ahmed: I love that. Follow your curiosity and just build something.
Marina, where can people learn about your work and find you online? 207 - > Marina Petzel: Sure, I share my LinkedIn posts constantly. Again, they are mostly related to AI for productivity, and you can find me on LinkedIn. I also have my website, marina.
petzel.tech, 208 - > Marina Petzel: where I share something about me, my talks, 209 - > Marina Petzel: and conferences that I am attending. I also have my Instagram, which I am trying to keep up with. In the path of AI it is really hard to keep up with everything, but I am trying to do my best to make sure that everyone stays connected and is able to learn more about AI.
210 - > Dewan Ahmed: We will be sure to link all of those in the podcast and video descriptions. That is Marina Petzel, Senior ML Engineer and AI Productivity Lead at Autodesk. 211 - > Dewan Ahmed: This was Ship Talk Podcast Season 4, Episode 5, and hey, if you are visiting AWS re:Invent this week, come say hello at Harness Booth 731. 212 - > Dewan Ahmed: Learn about the features we are shipping, and also grab some swag while you are there.
Thank you so much, Marina. 213 - > Marina Petzel: Thank you, Dewan, it was a pleasure to be here. 214 - > Dewan Ahmed: Likewise.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
The AI-Native Law Firm, with Ryan Walker of General LegalMeeting of the Minds · on Claude88 / 100
Agentic AI in the Commercial Workflow with Johnson Controls CDIO Vijay SankaranEnterprise AI Innovators · on LLMs (Large Language Models)88 / 100
Hiring top-tier talent, leveraging open source models, and staying competitive in the age of AI w/ Benny Chen #267The Engineering Leadership Podcast · on Reinforcement learning86 / 100
How SSW turned AI into ½ their pipeline - Ulysses Maclaren, COO of SSWSaaS Stories · on Claude86 / 100
Less about Models; More about ArchitecturePractical AI · on Computer vision85 / 100
Drones Are Changing the Future of FarmingFood Tech Talk · on Computer vision81 / 100ShipTalk Season 4 Finale: Engineering Excellence at AWS re:Invent67 / 100Crown Jewels In, Crown Jewels Out - The Hidden Risk of AI with Devan Shah (IBM)82 / 100
CTO Predictions for 2026: How AI Will Change Software Development (with Harness Field CTO Nick Durkin)78 / 100Beyond the Magic Box: Solving AI Hallucinations with Precision RAG (with Evgeny Ilinykh)85 / 100
Beyond Dashboards: How AI Is Redefining Developer Productivity with Adeeb Valiulla85 / 100