Company-Wide GenAI Transformation Through Hackathon-Driven Culture and Centralized Infrastructure

agoda · E-commerce · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.5 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise-wide GenAI adoption and production governance

Agoda needed to scale GenAI from hackathon experiments into hundreds of internal and customer-facing applications while controlling governance, cost, provider access, duplicated effort, and production quality.

Constraints:

Architecture

Agoda combined a builder-first culture with centralized GenAI infrastructure. A 2023 hackathon and the mastering-gpt community created demand and shared practice, while the GenAI Proxy centralized model access, legal/compliance review, routing, usage tracking, cost attribution, rate limits, and monitoring. An internal Chat Assistant Platform provided a secure PWA interface, custom assistants with pluggable tools, prompt/function playgrounds, and a Chrome extension, enabling more than 200 applications across support, engineering, marketing, customer-facing Q&A, reviews, and content curation.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  agoda-teams["Agoda teams"] --> gpt-hackathon["GPT hackathon"]
  gpt-hackathon --> mastering-gpt-community["mastering-gpt community"]
  mastering-gpt-community --> agoda-teams
  agoda-teams --> agoda-service-catalog["Agoda Service Catalog"]
  agoda-service-catalog --> compliance-and-legal-review["compliance and legal review"]
  compliance-and-legal-review --> genai-proxy["GenAI Proxy"]
  genai-proxy --> model-providers["model providers"]
  genai-proxy --> internal-chat-assistant-platform["Internal Chat Assistant Platform"]
  internal-chat-assistant-platform --> custom-assistant-tools["custom assistant tools"]
  internal-chat-assistant-platform --> chrome-extension["Chrome extension"]
  agoda-teams --> internal-chat-assistant-platform
  agoda-teams --> productivity-and-engineering-assistants["productivity and engineering assistants"]
  genai-proxy --> productivity-and-engineering-assistants
  custom-assistant-tools --> productivity-and-engineering-assistants
  genai-proxy --> customer-facing-genai-features["customer-facing GenAI features"]
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Agoda teams Developers, product teams, support, marketing, legal, finance, and other functions building or using GenAI applications. mastering-gpt community, Agoda Service Catalog, GenAI Proxy, Internal Chat Assistant Platform, productivity and engineering assistants
GPT hackathon Initial company-wide event that generated prototypes and momentum. Agoda teams, mastering-gpt community
mastering-gpt community Slack-based learning hub for prompt techniques, announcements, help, and shared GenAI practice. Agoda teams, Internal Chat Assistant Platform, productivity and engineering assistants
Agoda Service Catalog Registration point for GenAI use cases that triggers compliance and legal review. Agoda teams, compliance and legal review, GenAI Proxy
compliance and legal review Governance review for new use cases before GenAI access is granted. Agoda Service Catalog, GenAI Proxy
GenAI Proxy Central gateway for all GenAI traffic, provider routing, controls, usage tracking, cost attribution, rate limiting, and monitoring. Agoda Service Catalog, compliance and legal review, model providers, Internal Chat Assistant Platform, productivity and engineering assistants, customer-facing GenAI features
model providers External and Azure/OpenAI model backends routed through the proxy. GenAI Proxy
Internal Chat Assistant Platform Secure internal PWA with chat UI, custom assistants, pluggable tools, playground, and browser extension. Agoda teams, GenAI Proxy, custom assistant tools, Chrome extension
custom assistant tools Pluggable integrations for tasks such as SEO, prompt testing, data queries, and internal workflows. Internal Chat Assistant Platform, productivity and engineering assistants
Chrome extension Side chat for saved assistants that uses full page context. Internal Chat Assistant Platform
productivity and engineering assistants Internal tools such as Query Assist, AskGoda, Meeting Helper, coding agents, code review, documentation, monitoring, and MCP integrations. Agoda teams, GenAI Proxy, custom assistant tools
customer-facing GenAI features Traveler-facing Q&A assistants, review helpers, and content-curation systems built on the internal foundations. GenAI Proxy

Data flow

Teams register GenAI use cases through the Agoda Service Catalog and pass compliance/legal review. Approved applications access models through the GenAI Proxy, which routes requests across providers, tracks request/token usage and cost by service/team, applies rate limits and risk controls, and monitors usage. Employees use internal chat, Slack, Copilot-style tools, coding agents, and specialized assistants; successful patterns mature into maintained internal or customer-facing applications.

3 supporting quotes
“As part of registering a new use case in the Agoda Service Catalog, teams undergo a compliance and legal review process.”article, stated
“The GenAI Proxy gave teams a central gateway for all GenAI traffic, with built-in controls to support responsible usage.”article, stated
“Once approved, they gain access to GenAI capabilities without needing to manage authentication, billing, or usage tracking.”article, stated

Why this architecture

Agoda built shared foundations before broad product rollout because company-wide GenAI needed safe access, governance, cost attribution, provider flexibility, and internal learning loops to avoid isolated, duplicated, or risky experiments.

3 supporting quotes
“To scale responsibly, we began building shared foundations that could support growth without friction. ## The GenAI Proxy A core part of that foundation was our GenAI Proxy, a centralized access point for all GenAI traffic inside Agoda. It allowed us to: *”article, stated
“manage authentication, billing, or usage tracking. This centralized model enabled rapid experimentation while maintaining governance. ## Our Internal Chat Assistant Platform At the same time, we introduced our internal Chat Assistant Platform (akin to ChatGPT UI Clone) — a cost-effective, secure”article, stated
“build GenAI products responsibly and effectively. The goal wasn’t just to use GenAI; it was to learn how to build with it. * Learn how to build GenAI-powered applications: from designing assistant behavior to handling tool integrations, from shaping good UX to knowing when GenAI”article, stated

Models used: OpenAI.com models · Azure OpenAI models · GitHub Copilot · Cursor · Claude Code · OpenAI Codex · GenAI providers routed by the proxy

Scale

Alternatives considered (roads not taken)

Top-down strategy-first rollout

Rejected because: Agoda says the journey started with a hackathon rather than a strategy deck, creating bottom-up confidence and momentum.

“Our GenAI journey began not with a strategy deck or a top-down directive, but with a hackathon. In February 2023, over 200 developers across product and tech came together for a three-day company-wide event to explore one question: _What could we build if we”article, stated

Ungoverned teams calling providers directly

Rejected because: Direct use would duplicate authentication, billing, tracking, governance, and provider-routing work.

“growth without friction. ## The GenAI Proxy A core part of that foundation was our GenAI Proxy, a centralized access point for all GenAI traffic inside Agoda. It allowed us to: * Centralize compliance and legal review for new use cases, ensuring secure and responsible access. *”article, stated
“compliance and legal review process. Once approved, they gain access to GenAI capabilities without needing to manage authentication, billing, or usage tracking. This centralized model enabled rapid experimentation while maintaining governance. ## Our Internal Chat Assistant Platform”article, stated

External GenAI chat interfaces

Rejected because: Agoda wanted internal behavior control, lower per-seat cost, and a unified provider-agnostic interface.

“Rather than relying on external interfaces, we built our own platform for three key reasons: to define assistant behaviors in a GenAI-agnostic and internally controlled way”article, stated
“to avoid costly per-seat pricing models, and to ensure a unified interface that could work across any model provider.”article, stated

Start primarily with customer-facing GenAI

Rejected because: Agoda chose an Inside-Out approach to learn skills and infrastructure internally before higher-stakes customer features.

“Inside-Out” approach** to improving internal workflows and developing the skills, patterns, and infrastructure needed to build GenAI products responsibly and effectively.”article, stated
“internal use gave us the space to learn and adapt without the high stakes of a customer environment. To support this momentum, especially as mastering-GPT turned into an active learning hub and the Slack bot became a go-to tool, we focused on practical”article, stated

Patterns

model-routing (infrastructure)

The GenAI Proxy routes requests across OpenAI, Azure OpenAI, and additional providers as they were added.

“Route requests between different GenAI Providers as needed — we started off with OpenAI.com and Azure OpenAI models”article, stated

cost-aware-routing (operations)

Central proxy tracking attributes token/request usage and cost back to teams and avoids costly per-seat chat pricing.

“Track usage by service and team, down to the request and token level.”article, stated
“* Automatically attribute cost and usage back to teams. * Enforce rate limits, block risky calls, and monitor usage patterns.”article, stated

safety-policy-enforcement (governance)

Use cases receive compliance/legal review and proxy controls enforce rate limits, block risky calls, and monitor usage.

“Centralize compliance and legal review for new use cases, ensuring secure and responsible access.”article, stated
“Enforce rate limits, block risky calls, and monitor usage patterns.”article, stated

tool-use-orchestration (assistant-platform)

The internal assistant platform supports custom assistants with pluggable tools for SEO, prompt testing, data queries, and internal systems.

“Ability to create custom Assistants with pluggable tool support tailored for tasks like SEO, prompt testing, or data queries.”article, stated

eval-driven-development (quality)

Agoda explicitly treats testing, monitoring, and refinement as necessary skills for production GenAI rather than one-off demos.

“Figure out how to test GenAI features: what does quality look like for an assistant? How do you detect regressions? When do you simulate vs go live?”article, stated
“learning how to test, monitor, and refine GenAI applications at scale.”article, stated

human-in-the-loop-gate (culture)

The hackathon and mastering-gpt community form a human learning loop for sharing techniques, feedback, and production lessons.

“mastering-gpt quickly became the nerve center of our GenAI journey.”article, stated
“It wasn’t owned by any one team, and that was the point. It represented a self-sustaining community of people curious about what GenAI could do and eager to help each other figure it out.”article, stated

observability-tracing (operations)

The proxy tracks usage and monitors request patterns centrally across services and teams.

“Track usage by service and team, down to the request and token level.”article, stated
“level. * Automatically attribute cost and usage back to teams. * Enforce rate limits, block risky calls, and monitor usage patterns. Press enter or click to view image in full size GenAI Proxy allows Agoda Applications to Access Multiple Models across different Providers The GenAI Proxy”article, stated

Tradeoffs

Start with internal workflows before customer-facing GenAI.

Chose Inside-Out adoption over Immediate customer-facing rollout.

Rationale: Internal use let teams build skills, infrastructure, and testing practices without the high stakes of customer environments.

“internal use gave us the space to learn and adapt without the high stakes of a customer environment.”article, stated
“While much of our early GenAI focus was internal, the experience we gained in building robust GenAI systems laid a strong foundation for customer-facing innovation”article, stated

Centralize GenAI access through a proxy.

Chose GenAI Proxy over Each service managing provider authentication, billing, and tracking.

Rationale: Centralized access enables rapid experimentation while keeping governance, cost attribution, and monitoring consistent.

“Once approved, they gain access to GenAI capabilities without needing to manage authentication, billing, or usage tracking.”article, stated
“This centralized model enabled rapid experimentation while maintaining governance.”article, stated

Build an internal ChatGPT-like platform.

Chose Secure internal PWA assistant platform over Relying on external GenAI interfaces.

Rationale: Agoda wanted internal behavior control, lower pricing exposure, and a unified interface across providers.

“we built our own platform for three key reasons: to define assistant behaviors in a GenAI-agnostic and internally controlled way”article, stated
“to avoid costly per-seat pricing models, and to ensure a unified interface that could work across any model provider.”article, stated

Use grassroots community as the adoption engine.

Chose Hackathon plus mastering-gpt Slack community over Mandated top-down adoption only.

Rationale: The community created self-sustaining learning, project sharing, and curiosity that made adoption stick.

“A wave of GenAI curiosity swept through Agoda, setting the tone for everything that would follow — a grassroots, builder-first approach to GenAI.”article, stated
“It represented a self-sustaining community of people curious about what GenAI could do and eager to help each other figure it out.”article, stated

Gotchas (what bit them)

Internal context and data access are the main unlock for useful GenAI.

Root cause: The hackathon showed that raw model access was less valuable without the right internal context and data.

Resolution: Build internal platforms, tool integrations, MCPs, and assistant systems that expose company context safely.

“Access to internal context and the right data was the biggest challenge — and the key to unlocking value.”article, stated
“Custom MCPs: exposing Agoda’s internal systems, like build tools, documentation, and environments, to enrich coding assistants with additional context.”article, stated

Scaling GenAI creates governance, cost, and duplication problems.

Root cause: As tools, teams, and experiments multiply, direct access can fragment controls and duplicate work.

Resolution: Centralize GenAI access through a proxy with compliance review, provider routing, token/request tracking, cost attribution, rate limits, and monitoring.

“As interest in GenAI grew across the company, so did the number of tools, teams, and experiments.”article, stated
“How do we ensure good governance? How do we manage costs? How do we prevent duplicated effort?”article, stated

Prototype demos are much easier than reliable production applications.

Root cause: Production GenAI requires prompt/context engineering, testing, monitoring, and refinement at scale.

Resolution: Invest in infrastructure, safe deployment patterns, and skills that help teams bridge the prototype-to-production gap.

“Creating a GenAI-powered demo is deceptively easy, but building a high-quality, reliable production application is much harder.”article, stated
“developing skills in prompt and context engineering and learning how to test, monitor, and refine GenAI applications at scale.”article, stated

External GenAI interfaces can create cost and control problems.

Root cause: Per-seat pricing and external behavior definitions do not fit enterprise-wide internal usage and governance.

Resolution: Build an internal PWA assistant platform with provider-agnostic behavior definitions and unified access.

“Rather than relying on external interfaces, we built our own platform for three key reasons”article, stated
“to define assistant behaviors in a GenAI-agnostic and internally controlled way, to avoid costly per-seat pricing models”article, stated

Evaluation

Agoda evaluates the transformation through adoption metrics, application/service counts, ticket-automation rates, usage and token tracking through the proxy, and production practices around testing, monitoring, and refinement.

Metrics: 73% employee use of GenAI productivity tools · 45% of internal tech support tickets handled by GenAI on average · more than 200 applications and services by mid-2025 · over 200 hackathon developers · more than 40 hackathon ideas prototyped · 119 GenAI applications/services in 2023 · 204 GenAI applications/services by mid-2025 · request-level and token-level usage tracking · team-level cost attribution

4 supporting quotes
“applications and services, spanning everything from customer service and engineering to marketing and internal tools. And the impact is real, by the numbers (as of July 2025): * 73% of employees use GenAI productivity tools * On average, 45% of internal tech support tickets are handled by GenAI ”article, stated
“ 73% of employees use GenAI productivity tools * On average, 45% of internal tech support tickets are handled by GenAI * GenAI-driven features boosting customer experience and engagement In this blog, we’ll take a retrospective look at Agoda’s GenAI journey:”article, stated
“On average, 45% of internal tech support tickets are handled by GenAI”article, stated
“Track usage by service and team, down to the request and token level.”article, stated

Outcomes

The article reports a February 2023 hackathon with 200+ developers and 40+ ideas, growth from 119 GenAI apps/services in 2023 to 204 by mid-2025, 73% employee use of productivity tools, and 45% average automation of internal tech support tickets.

4 supporting quotes
“What began with a few experiments has grown into more than 200 applications and services”article, stated
“By Q1 2024, a third of our employees were using GenAI tools weekly. That number grew steadily, surpassing 50% by Q3 2024 and reaching 73% by Q2 2025.”article, stated
“Q&A Assistants: embedded across Agoda to help users get quick, natural language answers”article, stated
“These tools weren’t built as one-off demos; they’re maintained, used, and continuously improved as part of everyday workflows.”article, stated