
Sign up to save your podcasts
Or


How do you build a deepfake detector that journalists can trust? Danijela Horak, who led an AI research team at BBC Research and Development (R&D) at the time of recording, explains how her team approached deepfake detection, language models and responsible AI at a news organisation where trust comes first.
Danijela holds a PhD in pure mathematics and did two postdocs in applied maths before moving into machine learning in industry, then finance and the BBC. Christopher explores how that training shapes her work, from mastering fundamentals and setting clear hypotheses to careful baselining, and why she thinks the shift from training models to AI engineering is changing the role of data scientists. At BBC R&D, her team of around 25 computer vision and natural language processing scientists and engineers worked on projects from image restoration to automatic fact verification.
IN THIS EPISODE
• How the team fine tuned a smaller language model to rewrite Local Democracy Reporting Service stories in BBC style, which commercial models failed to capture.
• Why a classifier that she says matches or beats most commercial models still needs uncertainty quantification and interpretability before newsroom use.
• Why she treats deepfake detection for media as a high risk use case, closer to medicine than to detection at social media scale.
• How C2PA (Coalition for Content Provenance and Authenticity) metadata could change what we trust online.
• Why she believes responsible AI depends on transparency from model owners, given how little post training can change.
• Why she would never use language models for brainstorming, and where they do speed up learning.
"I think this worldview where you think everything is word bound and can be kind of put into words, it's completely wrong."
Danijela on the idea that you can prompt a model to do anything.
CHAPTERS
00:00 From pure maths to the BBC
03:35 Fundamentals, scientific thinking and the shift to AI engineering
11:12 Career advice, frontier research and the limits of prompting
16:13 Inside the BBC R&D AI team
19:00 Fine tuning a model to write in BBC style
21:18 How bad actors are using deepfakes
23:50 Building a deepfake detector journalists can trust
28:43 Can detection keep up? Watermarking and C2PA
30:39 Responsible AI and balancing innovation with trust
36:28 Leadership, stakeholders and predicting the future
44:35 Reinforcement learning, open endedness and using AI tools well
50:25 Critical thinking in the age of chatbots, and thanks
GUEST AND RESOURCES
Danijela Horak PhD
Mentioned in this episode:
• The Art of Doing Science and Engineering, Richard Hamming
• C2PA, the content provenance standard the BBC is working on with Microsoft, Google and Adobe
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 24 July 2025.
#DeepfakeDetection #ResponsibleAI #AIInJournalism #AI #Deepfakes #MachineLearning #ComputerVision #LLMs #Misinformation #ContentProvenance #C2PA #UncertaintyQuantification #ExplainableAI #AIEthics #ReinforcementLearning
How do you take generative AI and machine learning from proof of concept to production inside a large automotive company? Pavan Vemuri, Director of Product Engineering at SDVerse and formerly of Ford, explains how he scaled an internal ML platform and won over sceptical teams.
Christopher and Pavan trace his path from Cognizant and Stellantis to Ford, where he built MLOps (machine learning operations) tooling and became AI product and strategy lead for Ford's machine learning platform. Pavan explains why, in his experience, the hardest hurdles were organisational rather than technical, why in car assistants need careful governance, and how he now applies generative AI at SDVerse, a marketplace for automotive software.
IN THIS EPISODE
• Why the last 20% of a generative AI product, its guardrails and edge cases, decides whether it reaches production.
• Why his team chose Qdrant for a retrieval augmented generation (RAG) project, then settled on Elasticsearch to fit Ford's setup.
• How ML cleaned vehicle data at the source, catching anomalies such as trips that appear to last thousands of days.
• How automation cut onboarding for data scientists from five days to two hours.
• How early adopters, reusable components and adoption metrics took the platform to around 175 internal teams, by Pavan's account.
• Where he expects change next: AI agents, multimodality and edge AI.
"We can build the most sophisticated AI application in the world, if people do not want to use it, it fails." Pavan Vemuri
CHAPTERS
00:00 Welcome and Pavan's path from Cognizant to SDVerse
12:08 Career principles and lessons for AI product leaders
22:40 Future proofing yourself as an AI product manager
26:23 Going zero to one with generative AI at Ford
33:08 Generative AI in the vehicle and beyond
40:25 Why most generative AI projects never reach production
44:43 Innovation versus stability in large enterprises
48:20 MLOps in automotive: vehicle data and warranty issues
52:48 Cutting onboarding from five days to two hours
57:32 Scaling an ML platform across product teams
1:02:23 Generative AI in an automotive software marketplace
1:07:08 Building trust with engineering teams
1:11:28 What comes next: agents, multimodality and edge AI
1:16:06 Pavan's AI toolkit and a podcast recommendation
GUEST AND RESOURCES
Pavan Vemuri, Director of Product Engineering at SDVerse at the time of recording
SDVerse: https://www.sdverse.auto
Mentioned in this episode
• Qdrant, Elasticsearch, Kubeflow, Apache Airflow and Google BigQuery
• Claude and MCP (Model Context Protocol) servers
• Prompt Master Studio, Pavan's custom GPT
• Perplexity and the Perplexity API
• Lex Fridman's podcast conversation with Andrej Karpathy
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 25 July 2025.
#GenerativeAI #AutomotiveAI #MLOps #AI #MachineLearning #AIProductManagement #SoftwareDefinedVehicles #EnterpriseAI #AIAdoption #MLPlatform #RAG #SemanticSearch #AIAgents #EdgeAI #ProductLeadership
How do you build AI agents that learn what each individual user of a consumer app wants? Schaun Wheeler PhD, cofounder and Chief Scientist of Aampe at the time of recording, explains why his team assigns a dedicated learning agent to every user, and why those agents do not depend on large language models (LLMs).
In this episode of The NeuralPod, Christopher talks with Schaun about his route from anthropology fieldwork in Kyrgyzstan to machine learning (ML) for the US Army, and why that background convinces him most ML and agent projects fail on the human problem, not the model.
He describes Aampe's aim as adapting the app experience to each user, rather than making users fit the business, and why A/B tests, bandits and standard ML were not enough.
IN THIS EPISODE
• Why Schaun sees data science as an unkept promise, citing survey estimates he recalls of 70 to 90% of ML projects never yielding value.
• Why LLMs suit kind learning environments with stable rules, while customer engagement is a wicked one that needs semantic memory.
• How each agent reads the full app event stream, adapts difference in differences to measure impact against business as usual, and stores results as beta distributions.
• How agents shift between exploring and exploiting when a user's life changes.
• Why Schaun treats LLM coding assistants as a pair coder, not a replacement for engineers.
• How Aampe runs a fully remote team on trust, blameless retrospectives and working trips.
On blameless retrospectives, Schaun says: "You can't not do stupid things when you're building a startup 'cause you're moving so fast."
CHAPTERS
00:00 Welcome: from cognitive anthropology to data science
04:13 The unkept promise of data science and agentic AI
07:31 Why ML and agent projects fail: misreading the human problem
09:27 Navigating a data science career when employers are unsure
12:11 Will coding assistants replace software engineers?
15:12 What Aampe does: a dedicated agent for every user
19:49 Why Aampe does not rely on LLMs
27:35 How each agent learns from a single user
34:34 Judging agents like a money manager
37:34 Building Aampe: rethinking content tools for agents
40:41 Running a fully remote team on trust and ownership
49:19 Where agents are heading next
50:42 Staying intentional with AI tools, and a book on evidence
55:20 What excites Schaun about AI
GUEST AND RESOURCES
Aampe: https://aampe.com
The Book of Evidence by Peter Achinstein
Robin Hogarth on kind and wicked learning environments
Tools Schaun uses: ChatGPT and Gemini
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 8 July 2025.
#AIAgents #AgenticAI #AI #CustomerEngagement #ConsumerApps #Personalisation #DataScience #MachineLearning #CausalInference #Experimentation #ABTesting #LifecycleMarketing #RemoteWork #StartupCulture #MarTech
How do you personalise a retail homepage at scale, and what does it take to get recommender systems into production? Shreesha Jagadeesh, Associate Director of Applied Machine Learning at Best Buy at the time of recording, shares practical lessons on ranking, experimentation and leading machine learning (ML) teams.
In this episode of The NeuralPod, Christopher talks with Shreesha about a path from hardware engineering at a biomedical startup, through Cisco, Ernst & Young and Amazon, to Best Buy in mid 2022. At its centre is Shreesha's industry paper for the ACM Conference on Recommender Systems (RecSys), in which an XGBoost model ranks the themed stories and nested carousels on Best Buy's homepage. Shreesha speaks in a personal capacity, not as a Best Buy spokesperson.
IN THIS EPISODE
• Why retail recommendations differ from social feeds, including a milder cold start problem and fewer signals for logged out visitors.
• Launching the homepage model before any training data existed, and keeping latency low with parallel feature store calls and multi region deployment.
• Why a two week A/B test can miss the long term effects of a loyalty offer, and how a test without any model can set a benchmark.
• What Shreesha looks for when hiring, including online deployment experience and an understanding of train serve skew.
Shreesha reports a large increase in homepage engagement, overall latency well within 100 milliseconds and gains in an internal revenue metric.
On hiring, Shreesha notes: "The online serving use case is a whole another ball game compared to any sort of batch deployment."
CHAPTERS
00:00 Welcome and Shreesha's path into machine learning
11:02 Joining Best Buy: retail versus HR tech recommendations
15:07 Transferable ML skills and early career advice
23:21 Retail challenges: cold start and logged out visitors
28:28 The RecSys paper: ranking homepage content with XGBoost
33:27 Launching without training data, latency and results
39:16 Getting buy in and designing A/B tests
46:00 Representation learning and LLM based recommenders
50:50 Leadership and why more ML projects reach production
57:59 Hiring ML engineers and researchers
1:02:41 An end to end view of enterprise recommender systems
1:10:11 AI tools: Copilot and ChatGPT Deep Research
1:13:03 Where recommender systems go next: LLM cost and latency
1:16:19 Closing thoughts
GUEST AND RESOURCES
Shreesha Jagadeesh, Associate Director of Applied Machine Learning at Best Buy (at the time of recording)
Mentioned in this episode:
• Shreesha's ACM RecSys 2025 industry paper on homepage personalisation with XGBoost
• Hidden Technical Debt in Machine Learning Systems (Google, 2015)
• HSTU, a transformer based recommendation paper from Meta
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 17 May 2025.
#RecommenderSystems #RecSys #AI #MachineLearning #Personalisation #MLEngineering #RetailAI #Ecommerce #XGBoost #ABTesting #FeatureStore #RepresentationLearning #LLMs #MLOps #AILeadership
What is an AI agent, and how should a business adopt AI so that it pays off? Erik Schwartz, Chief AI Officer at Tricon at the time of recording, answers both with a practical view shaped by 30 years in search and information retrieval.
In this episode of The NeuralPod, Christopher talks with Erik about a career that runs from digital libraries at the Naval Research Laboratory in the mid 1990s, through enterprise search at Convera, Fast and Microsoft, to Comcast and then Elsevier, where Erik's team built Scopus AI on a trusted database of over 90 million research abstracts.
IN THIS EPISODE
• Erik's definition of an agent: a system that takes a goal, builds and refines a plan, uses tools and draws on long and short term memory.
• Why simple automation with tools such as Make or Zapier often solves the problem before an agent is needed, and why AI literacy and clean data matter.
• How a safe sandbox and keeping humans in the loop help AI move from proof of concept to production.
• Measuring return on investment (ROI) through leading indicators such as new revenue, and trailing ones such as time saved.
• What agents could mean for search engine optimisation (SEO), advertising and enterprise search, which Erik believes finally has a real path forward.
• Erik's reading of DeepSeek: training on complete plans, working within constrained compute and releasing it as open source.
On agents, Erik puts it simply: "It's really intelligent workflow with the ability to think as it moves."
CHAPTERS
00:00 Welcome and the pace of change in AI
02:25 Erik's career, from digital libraries to Scopus AI
08:36 Dotcom lessons and the wrapper app debate
13:48 What is an AI agent? Goals, plans, tools and memory
18:00 Adopting agents: why simple automation often comes first
22:39 Planning an AI roadmap: literacy, ROI and model costs
25:58 The chief AI officer role and barriers to adoption
33:26 Prototype to production and measuring ROI
40:06 Finding product market fit faster with AI
43:05 How agents change SEO, websites and enterprise search
49:54 Agents, ad revenue and distributed ledgers
55:17 DeepSeek: reasoning, training on plans and constrained compute
1:01:47 What excites Erik: multimodality, robotics and unified assistants
1:05:23 Closing thoughts and how to connect with Erik
GUEST AND RESOURCES
Erik Schwartz, Chief AI Officer at Tricon (at the time of recording)
Mentioned in this episode: OpenAI Deep Research and Operator, DeepSeek, Make, Zapier, Lovable
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 13 February 2025.
#AIAgents #EnterpriseAI #AI #AgenticAI #AIAdoption #ChiefAIOfficer #AILiteracy #GenerativeAI #LLMs #EnterpriseSearch #InformationRetrieval #DeepSeek #AIStrategy #Automation #ReasoningModels
When should an AI startup hire a chief of staff, and what should founders know before raising venture capital? Mwenya Kawesha, a chief of staff who has worked in UK AI startups, explains how to scope the role, when to bring one in and what she has learned from fundraising.
In this episode of The NeuralPod, Christopher talks with Mwenya about her route from studying history at Oxford, via the Clinton Foundation and Accenture's business strategy team, into AI startups. There she led technical delivery teams, focused on using AI to tackle harmful content online, and moved through business operations into chief of staff roles.
IN THIS EPISODE
• Why the role must fit the founder and leadership team, from copilot to a sole founder to a narrower role in a larger team.
• Why Mwenya rates a willingness to learn above prior technical experience, while valuing fluency in the technical team's language.
• The signs a founder needs support, and why she suggests hiring possibly a year before a big round.
• Treating AI strategy as part of business strategy: evidencing the problem, checking the data and sometimes asking whether SQL would do the job.
• Balancing regulation and speed across the EU AI Act, US federal policy and California, and why responsible AI matters to large clients.
• Fundraising lessons: knowing what you are raising and why, choosing between equity, debt and convertible loan notes, and why she sees UK investors expecting an earlier path to profitability.
Mwenya's rule of thumb: "Basically hire a chief of staff before everyone is telling you, why don't you have a chief of staff?"
CHAPTERS
00:00 Welcome and introducing Mwenya
00:45 From history at Oxford to Accenture and AI startups
06:02 What does a good chief of staff do?
10:56 Does a chief of staff need technical knowledge?
13:53 Responsible AI development and who to involve
16:56 Building an AI strategy and testing your assumptions
22:53 Why AI teams need diverse skills and good storytelling
25:57 Regulation in the EU and US: balancing safety and speed
29:54 When should a founder hire a chief of staff?
33:10 Why responsible AI matters for winning clients
35:35 Where founders should start with fundraising
40:20 Raising in the UK versus the US in 2025
42:38 The biggest challenges of running a raise
46:09 What investors look for at each stage
GUEST AND RESOURCES
Guest: Mwenya Kawesha, Chief of Staff
Book mentioned: Venture Deals by Brad Feld and Jason Mendelson
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 31 January 2025.
#ChiefOfStaff #AIStartups #RaisingVC #AI #VentureCapital #StartupFundraising #ResponsibleAI #AIStrategy #AIRegulation #EUAIAct #StartupOperations #FounderAdvice #AIEthics #StartupHiring #TechnicalFounders
How do you evaluate large language models (LLMs) and AI agents once simple accuracy scores stop telling the story? Marco Del Tredici, a natural language processing (NLP) researcher who has worked at Amazon Alexa and Cohere, explains how evaluation, interpretability and agent research are changing.
In this episode of The NeuralPod, Christopher talks with Marco about his route from formal linguistics to a PhD at the University of Amsterdam, where he studied how knowing about social media users helps models understand their language. He then worked on multi turn conversations at Amazon Alexa, and on retrieval augmented generation (RAG) and agents at Cohere. At the time of recording he was working on autonomous agents at Twin.
He also covers where agents can help most, why multi agent setups look most realistic for now, and the shift he expects towards people and agents working on tasks together.
IN THIS EPISODE
• Why Marco sees signs that scaling data and compute alone may be reaching a plateau, and why more reasoning at inference time is one response.
• How evaluation moved beyond F1, accuracy, precision and recall, and why faithfulness to retrieved documents and low inter annotator agreement make it harder.
• Why he favours task specific metrics, such as whether an agent managed to book a flight, and the trade off in broader understanding.
• Two routes to interpretability: inspecting a model's internal activations, and asking it to explain its steps through chain of thought and self critique.
• His view of agents as a scale of behaviours: planning and replanning, tool use, memory and human in the loop collaboration.
• Open problems for autonomous agents: missing trajectory data in pretraining, knowing when to backtrack and building longer term memory.
CHAPTERS
00:00 Introduction
01:01 From linguistics to NLP research, Amazon Alexa, Cohere and Twin
06:18 Moving from a PhD into industry research
09:47 Are LLMs hitting a plateau? Inference time and transformers
15:08 Why evaluating LLMs keeps getting harder
21:49 Is there a gold standard for evaluating LLMs?
23:42 Interpretability: activations and chain of thought
28:50 What is an AI agent?
37:12 Research challenges for autonomous agents
43:51 Should agents learn like humans? Lessons from AlphaGo
46:18 Where agents can make a real world impact
49:25 Many specialised agents or one assistant?
52:09 Predictions: people and agents working together
54:54 Closing thanks
GUEST AND RESOURCES
Guest: Marco Del Tredici PhD
Mentioned in this episode: OpenHands (formerly OpenDevin), AlphaGo, the MMLU benchmark, and the BLEU and ROUGE metrics.
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
#LLMEvaluation #AIAgents #Interpretability #AI #LLMs #AutonomousAgents #AgenticAI #NLP #RAG #ChainOfThought #MachineLearning #AIResearch #MultiAgentSystems #HumanInTheLoop #Transformers
How do you build search and recommender systems (RecSys) for a marketplace where every home is unique and hundreds of people may want the same one? Tetiana Torovets, Head of AI at QuintoAndar, explains how she ranks limited supply fairly, measures long term impact and builds an AI search assistant for property.
In this episode of The NeuralPod, Christopher and Tetiana discuss ranking and personalisation for one of a kind inventory, drawing on her earlier work at ThredUp, the secondhand fashion marketplace, and at QuintoAndar, a property technology company. Her background spans applied mathematics, banking and fintech consulting.
She also explains the AI search assistant she presented at the Haystack conference in Berlin, which pairs a large language model (LLM) chatbot that turns conversation into search filters with visual language models that search images. The aim is to close the gap between a site full of filters and a good estate agent who understands your context.
IN THIS EPISODE
• Why unique, limited supply makes fairness and the distribution of demand a core business problem.
• How ThredUp priced unique secondhand items using image, brand and sales history data, and linked A/B tests to profit.
• Moving from a system of ensemble tree models, with candidate generation, ranking and reranking, towards multi objective neural networks.
• Long term holdout groups, funnel and proxy metrics, and why Tetiana has repeatedly seen companies build their own A/B testing tools.
• Why users often treat an AI assistant like a search box, and how voice input and onboarding suggestions can help.
• Why, in cities such as Lisbon, San Francisco and Rio de Janeiro, the best alternative neighbourhood is not always the closest.
CHAPTERS
00:00 Introducing Tetiana Torovets
01:31 From applied mathematics and fintech to search
07:04 Marketplaces where every item is unique
12:42 Pricing unique items and tying tests to profit
18:03 Multi objective ranking and the move to neural networks
20:48 Cold start for new users and new listings
24:01 Winning stakeholder buy in with long term holdouts
29:44 A/B testing lessons: contamination, proxy metrics and in house tools
34:19 Search, recommendations, embeddings and LLMs
39:06 Building an AI search assistant for property
43:55 Could AI replace estate agents? Lazy users and voice
48:06 Where property tech is heading: location and personalisation
54:16 Hiring ML engineers and leading a team
57:54 Closing thoughts
GUEST AND RESOURCES
QuintoAndar: https://www.quintoandar.com.br
ThredUp: https://www.thredup.com
Mentioned: Tetiana's Haystack talk in Berlin on QuintoAndar's AI search assistant, and the RecSys conference.
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 18 October 2024.
#RecommenderSystems #AISearch #PropTech #AI #MachineLearning #Personalisation #RecSys #SearchRanking #ConversationalSearch #ABTesting #Marketplaces #LLMs #GenerativeAI #ColdStart #DataScience
How do travel platforms price and recommend hotels when most people only book a few trips a year? Jan Krasnodebski, Director of AI at Hostelworld when this episode was released, draws on around 13 years at Expedia to explain pricing strategy, personalisation and recommender systems (RecSys).
In this episode of The NeuralPod, Christopher and Jan begin with pricing: the leverage Jan sees in it, and why he argues it must support the business and marketing strategy. On generative AI, Jan sees real value in customer service, but argues that large language models (LLMs) are weak at maths and lack the live pricing and availability that good hotel recommendations depend on.
Later, they cover recommendations that serve several stakeholders at once, the ethics of using traveller data, and what Jan values when leading and hiring in AI: curiosity, humility and the ability to explain your work.
IN THIS EPISODE
• Why Jan believes relaunching a heart valve at five times its original price helped a whole industry grow.
• Jan's case for layering generative AI on top of foundational recommendation models, adding explanations and a feedback loop for travellers.
• Why Jan calls travel personalisation a continuous cold start problem, citing figures of around 60 days of research and, depending on the source, 30 to 40 site visits before booking.
• Balancing traveller, supplier and platform goals with meta learning, and internal priorities with multi armed bandits and a common currency.
• Handling sensitive signals such as LGBTQ friendly filters and accessibility needs, and why Jan sees most fairness questions on the supplier side.
• Why Jan asks some senior candidates for a presentation rather than a take home assessment.
As Jan puts it: "You can only price to the degree that you're providing value to a customer."
CHAPTERS
00:00 Introducing Jan Krasnodebski
01:04 Why pricing? Lessons from a heart valve
05:27 How should you price a new product?
07:04 Why LLMs struggle with pricing maths
08:49 Generative AI in travel and live prices
15:27 Personalisation as a continuous cold start problem
24:12 Incognito visitors and planning your data first
27:22 Weather, trip details and third party data
30:56 Conflicting signals and improving on a good baseline
33:05 Multi stakeholder recommendations, meta learning and bandits
39:56 Ethics, anonymisation and sensitive data
46:56 Leading in AI: curiosity and humility
51:32 How Jan evaluates researchers and engineers
1:00:50 Closing question: what's next for AI?
GUEST AND RESOURCES
Hostelworld: https://www.hostelworld.com
Mentioned: Expedia's first Kaggle competition (2013), the MovieLens dataset and the RecSys conference.
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
Recorded 6 June 2024.
#RecommenderSystems #DynamicPricing #TravelTech #AI #MachineLearning #Personalisation #RecSys #GenerativeAI #PricingStrategy #MultiArmedBandits #ResponsibleAI #ColdStart #OnlineTravel #AILeadership #DataScience
What are digital twins, and why do they matter beyond engineering? Sumit Singh PhD, of Fujitsu, explains how digital twins bring data, simulation and the physical world together, and what organisations and individuals can do about the skills gap.
In this episode of The NeuralPod, Christopher talks with Sumit about his route from an MSc group project to a PhD on digital twins in aerospace. Sumit describes a digital twin as a virtual representation of a physical system or process, built on end to end integration of data, and explains why he doubts one can exist without simulation.
Christopher recalls mapping 150 to 200 people in the UK with the skills for an applied research role in digital twins, two or three years before the recording, and Sumit sets out how enterprises, universities and government could close that gap.
IN THIS EPISODE
• How Sumit sees digital twins exposing unknowns and interdependencies, such as digitally rehearsing transport policy changes before they reach the real world.
• How digital twins can help government prepare for emergencies such as flooding, and support a more proactive approach to investment.
• Why organisations hesitate to share data, and how connected digital twins aim to harmonise data sharing for collective value.
• Where to start: pick a stage of the product life cycle and assess the value before the engineering work begins.
• The challenges Sumit highlights, from value realisation and cultural change to crossover skills that pair data science with systems architecture and domain knowledge.
• His advice for newcomers: you do not need a PhD; start with one sector, watch the others and stay curious.
"I mean, I can think of digital twin for everything, but you've got to start with somewhere."
Sumit Singh
CHAPTERS
00:00 Show introduction
00:43 Meet Sumit Singh
01:34 What is a digital twin and why does it matter?
07:00 Where digital twins create value for customers
10:11 Digital twins for public good
13:37 Data, accuracy and connected digital twins
16:06 First steps for organisations
19:03 Key challenges: value, culture and crossover skills
21:35 Simulation, the metaverse and virtual training
27:33 Future skills and reskilling
33:16 Closing the digital twin skills gap
39:00 Is a PhD necessary? Advice for starting out
43:47 The ideal future: automation, interoperability and privacy
49:08 Closing thoughts
GUEST AND RESOURCES
Sumit Singh PhD, Fujitsu
Mentioned in the episode: the National Digital Twin Programme, and a 2022 BBC article Sumit recalls about a digital twin of every human by 2030.
FOLLOW THE NEURALPOD
Subscribe for conversations with the founders, researchers and engineers building and applying AI.
YouTube: https://www.youtube.com/@TheNeuralPod
Apple Podcasts: https://podcasts.apple.com/us/podcast/the-neuralpod/id1761461659
The NeuralPod is hosted by Christopher Coyne, founder of NeuralRec: https://neuralrec.ai
#DigitalTwins #Simulation #SkillsGap #AI #Metaverse #ConnectedDigitalTwins #DigitalEngineering #DigitalTransformation #Interoperability #DataSharing #Aerospace #Manufacturing #CriticalInfrastructure #Reskilling #MachineLearning