West Coast AI Engineering & Infrastructure Summit 2027

Beat the rush and join us early for complimentary barista-made coffee and breakfast.

In a market where foundation model releases and funding rounds move faster than most engineering roadmaps, this opening keynote examines why AI teams increasingly can't afford the "build fast, harden later" approach that worked in earlier tech cycles, and sets the tone for a day focused on building production-grade infrastructure at startup speed.

• How the compressed timeline between prototype, funding round, and real user traffic is leaving teams less runway than previous software cycles to retrofit infrastructure after initial success.
• Distinguishing between genuinely necessary early architectural decisions and premature optimisation that slows a team down without reducing real risk.
• What infrastructure investments AI teams are making early that are paying off, and where "moving fast" has quietly become the excuse for skipping work that will cost far more later.

With unmatched proximity to frontier AI research, engineering teams face a unique version of the production gap: promising research techniques arrive faster than anywhere else, but the discipline required to turn a paper's findings into a reliable production system hasn't kept pace with the excitement.

• Common patterns where techniques that looked compelling in a paper or benchmark broke down under the messier conditions of real production traffic and data.
• How some teams are building internal evaluation rigour to separate genuinely production-ready techniques from impressive demos, resisting pressure to ship the newest approach simply because it's newest.
• Concrete engineering lessons from teams who've successfully taken a research technique from paper to reliable production deployment, and what got cut or hardened along the way.

With new frontier and open-weight models shipping faster than most engineering teams can evaluate them, this panel brings together AI and ML platform leaders to discuss how they're making model selection and infrastructure decisions durable enough to survive the next model release cycle, rather than re-architecting every few months.

• How the constant temptation to swap in the latest, most capable model is creating hidden engineering and evaluation costs that rarely get accounted for in the decision.
• Practical architectural patterns that let teams change underlying models without rewriting downstream application logic every time.
• Under what genuine technical and cost conditions teams are finding fine-tuning smaller models still outperforms simply throwing a larger frontier model at the problem.

After a major foundation model provider outage left a consumer AI app completely unusable for six hours during peak traffic, one startup's infrastructure lead rebuilt the system around genuine multi-model fallback, without doubling engineering overhead or degrading the product experience during normal operation.

• How the team discovered that fallback logic they assumed existed was largely superficial, with core product functionality entirely dependent on one provider's availability.
• The specific architectural approach used to gracefully degrade functionality during a provider outage rather than building and maintaining a fully duplicated experience across models.
• How the team built regular failover drills into their engineering practice, rather than only discovering gaps when the next real outage hit.

A hands-on, interactive session working through a real production AI scenario as a room. Details announced soon.

Facing explosive user growth that threatened to overwhelm infrastructure built for an early-access cohort, one developer tools startup's platform lead scaled their AI coding assistant through targeted architectural changes rather than the ground-up rebuild the team initially assumed would be necessary.

• How the team used load testing and bottleneck analysis to isolate the specific components that would break at scale, avoiding a costly and unnecessary full rewrite.
• The specific caching, batching, and model-routing changes made to keep unit economics viable as usage scaled far faster than the original cost model anticipated.
• How the team balanced the urgency of the scaling work against the risk of shipping degraded product quality under pressure.

After silently shipping a model update that degraded output quality for weeks before anyone noticed, one autonomous systems company's ML platform lead built a continuous evaluation pipeline that caught quality regressions before they reached production.

• How the team traced the gap back to relying entirely on manual spot-checks and user complaints rather than systematic, continuous evaluation.
• The specific evaluation checkpoints introduced before any model update could reach production, catching quality issues that manual review had missed.
• How the platform lead made the case internally that adding evaluation gates, despite slowing release velocity slightly, was a net win given the cost of the regression that had gone undetected.

Financial services, healthcare, and other regulated industries are facing increasingly specific requirements around AI system auditability and explainability. This panel explores how engineering teams are building compliance directly into system architecture rather than bolting it on after the fact.

• Concrete examples of the evidence and documentation regulators expect when reviewing AI systems used in regulated contexts.
• Technical patterns for capturing decision provenance, model versioning, and prompt history in a way that satisfies compliance review without crippling system performance.
• Honest comparison of the effort and technical debt incurred by teams who built compliance requirements in from day one versus those who tried to add them after a system was already in production.

Small-group, discussion-based sessions where you'll work through real AI engineering and infrastructure challenges with peers in similar roles. Roundtable topics will be announced soon.

Put your knowledge to the test in this fast-paced quiz covering real-world trivia, key concepts, and emerging trends. Compete for bragging rights — and a voucher — as the top scorer takes the crown.

Facing exponential growth in user-generated content and mounting pressure to catch harmful content faster, one consumer social platform's infrastructure leader rebuilt their moderation pipeline to operate at real-time scale without the cost or latency blowing out as volume grew.

• How the team recognised that simply adding more compute to their existing moderation architecture would make costs untenable well before reaching their growth targets.
• The specific routing architecture built to apply lightweight, fast filtering to low-risk content while reserving deeper, more expensive model review for genuinely ambiguous or high-risk cases.
• How the team validated that the tiered approach didn't sacrifice moderation quality even as it cut the average processing cost per piece of content.

Foundation model providers are shipping new capabilities and deprecating old ones at a pace no other software dependency has ever moved at. This keynote examines how product and infrastructure teams are building resilient systems on top of a foundation they don't control and can't predict.

• How the pace and unpredictability of frontier model changes differs from any prior software dependency relationship, and why existing vendor management practices don't fully transfer.
• Practical architectural patterns teams are using to detect and absorb sudden shifts in model behaviour, pricing, or availability without a production outage.
• Real experience from teams who've built multi-model fallback strategies, and what that resilience has actually cost them in complexity and performance.

The demand for engineers who can build and operate production AI infrastructure now outpaces supply. This interactive session brings together engineering leaders to discuss what skills genuinely differentiate strong AI infrastructure talent, and how teams are building capability when the ideal hire rarely exists on the open market.

• What engineering leaders are finding actually correlates with success building and operating production AI infrastructure, versus credentials that look impressive on paper.
• How teams are weighing the cost and time of upskilling existing engineers against the increasingly expensive and competitive market for experienced AI infrastructure hires.
• Emerging patterns in how leading organisations are organising the roles and responsibilities across ML platform, infrastructure, and application engineering functions.

Unwind with your peers for a couple of drinks on us!

Beat the rush and join us early for complimentary barista-made coffee and breakfast.

In a market where foundation model releases and funding rounds move faster than most engineering roadmaps, this opening keynote examines why AI teams increasingly can't afford the "build fast, harden later" approach that worked in earlier tech cycles, and sets the tone for a day focused on building production-grade infrastructure at startup speed.

• How the compressed timeline between prototype, funding round, and real user traffic is leaving teams less runway than previous software cycles to retrofit infrastructure after initial success.
• Distinguishing between genuinely necessary early architectural decisions and premature optimisation that slows a team down without reducing real risk.
• What infrastructure investments AI teams are making early that are paying off, and where "moving fast" has quietly become the excuse for skipping work that will cost far more later.

With unmatched proximity to frontier AI research, engineering teams face a unique version of the production gap: promising research techniques arrive faster than anywhere else, but the discipline required to turn a paper's findings into a reliable production system hasn't kept pace with the excitement.

• Common patterns where techniques that looked compelling in a paper or benchmark broke down under the messier conditions of real production traffic and data.
• How some teams are building internal evaluation rigour to separate genuinely production-ready techniques from impressive demos, resisting pressure to ship the newest approach simply because it's newest.
• Concrete engineering lessons from teams who've successfully taken a research technique from paper to reliable production deployment, and what got cut or hardened along the way.

With new frontier and open-weight models shipping faster than most engineering teams can evaluate them, this panel brings together AI and ML platform leaders to discuss how they're making model selection and infrastructure decisions durable enough to survive the next model release cycle, rather than re-architecting every few months.

• How the constant temptation to swap in the latest, most capable model is creating hidden engineering and evaluation costs that rarely get accounted for in the decision.
• Practical architectural patterns that let teams change underlying models without rewriting downstream application logic every time.
• Under what genuine technical and cost conditions teams are finding fine-tuning smaller models still outperforms simply throwing a larger frontier model at the problem.

After a major foundation model provider outage left a consumer AI app completely unusable for six hours during peak traffic, one startup's infrastructure lead rebuilt the system around genuine multi-model fallback, without doubling engineering overhead or degrading the product experience during normal operation.

• How the team discovered that fallback logic they assumed existed was largely superficial, with core product functionality entirely dependent on one provider's availability.
• The specific architectural approach used to gracefully degrade functionality during a provider outage rather than building and maintaining a fully duplicated experience across models.
• How the team built regular failover drills into their engineering practice, rather than only discovering gaps when the next real outage hit.

A hands-on, interactive session working through a real production AI scenario as a room. Details announced soon.

Facing explosive user growth that threatened to overwhelm infrastructure built for an early-access cohort, one developer tools startup's platform lead scaled their AI coding assistant through targeted architectural changes rather than the ground-up rebuild the team initially assumed would be necessary.

• How the team used load testing and bottleneck analysis to isolate the specific components that would break at scale, avoiding a costly and unnecessary full rewrite.
• The specific caching, batching, and model-routing changes made to keep unit economics viable as usage scaled far faster than the original cost model anticipated.
• How the team balanced the urgency of the scaling work against the risk of shipping degraded product quality under pressure.

After silently shipping a model update that degraded output quality for weeks before anyone noticed, one autonomous systems company's ML platform lead built a continuous evaluation pipeline that caught quality regressions before they reached production.

• How the team traced the gap back to relying entirely on manual spot-checks and user complaints rather than systematic, continuous evaluation.
• The specific evaluation checkpoints introduced before any model update could reach production, catching quality issues that manual review had missed.
• How the platform lead made the case internally that adding evaluation gates, despite slowing release velocity slightly, was a net win given the cost of the regression that had gone undetected.

Financial services, healthcare, and other regulated industries are facing increasingly specific requirements around AI system auditability and explainability. This panel explores how engineering teams are building compliance directly into system architecture rather than bolting it on after the fact.

• Concrete examples of the evidence and documentation regulators expect when reviewing AI systems used in regulated contexts.
• Technical patterns for capturing decision provenance, model versioning, and prompt history in a way that satisfies compliance review without crippling system performance.
• Honest comparison of the effort and technical debt incurred by teams who built compliance requirements in from day one versus those who tried to add them after a system was already in production.

Small-group, discussion-based sessions where you'll work through real AI engineering and infrastructure challenges with peers in similar roles. Roundtable topics will be announced soon.

Put your knowledge to the test in this fast-paced quiz covering real-world trivia, key concepts, and emerging trends. Compete for bragging rights — and a voucher — as the top scorer takes the crown.

Facing exponential growth in user-generated content and mounting pressure to catch harmful content faster, one consumer social platform's infrastructure leader rebuilt their moderation pipeline to operate at real-time scale without the cost or latency blowing out as volume grew.

• How the team recognised that simply adding more compute to their existing moderation architecture would make costs untenable well before reaching their growth targets.
• The specific routing architecture built to apply lightweight, fast filtering to low-risk content while reserving deeper, more expensive model review for genuinely ambiguous or high-risk cases.
• How the team validated that the tiered approach didn't sacrifice moderation quality even as it cut the average processing cost per piece of content.

Foundation model providers are shipping new capabilities and deprecating old ones at a pace no other software dependency has ever moved at. This keynote examines how product and infrastructure teams are building resilient systems on top of a foundation they don't control and can't predict.

• How the pace and unpredictability of frontier model changes differs from any prior software dependency relationship, and why existing vendor management practices don't fully transfer.
• Practical architectural patterns teams are using to detect and absorb sudden shifts in model behaviour, pricing, or availability without a production outage.
• Real experience from teams who've built multi-model fallback strategies, and what that resilience has actually cost them in complexity and performance.

The demand for engineers who can build and operate production AI infrastructure now outpaces supply. This interactive session brings together engineering leaders to discuss what skills genuinely differentiate strong AI infrastructure talent, and how teams are building capability when the ideal hire rarely exists on the open market.

• What engineering leaders are finding actually correlates with success building and operating production AI infrastructure, versus credentials that look impressive on paper.
• How teams are weighing the cost and time of upskilling existing engineers against the increasingly expensive and competitive market for experienced AI infrastructure hires.
• Emerging patterns in how leading organisations are organising the roles and responsibilities across ML platform, infrastructure, and application engineering functions.

Unwind with your peers for a couple of drinks on us!