AdRoom
I took ownership of Pixis's AI creative platform from a previous agency and led its 6-engineer team, shipping UGC video, video scoring and credit billing.
- Client
- Pixis
- My role
- Engineering lead
- Team
- 6 engineers
- Period
- 2026
Results
6
Engineers led across 3 production reposTeam lead
~170
PRs reviewed and mergedin 4 months
28to 0
Known critical and high advisories, pixis.ai CMS−28
Context
AdRoom is Pixis’s AI ad-creative studio. Brands upload product images, and AdRoom generates scenes, full ads, copy and UGC-style video, then runs them through approval workflows. It is three production services (Node/Express, React and Python FastAPI) on Kubernetes in AWS.
In April 2026, Pixis moved AdRoom’s engineering away from a previous agency, and I took ownership. A month later, Pixis’s corporate site, pixis.ai, and its CMS came over too.
My role
Engineering lead. I led and managed a team of 6 engineers, was code owner and release gatekeeper for the 3 production repos, and was the engineering contact for Pixis’s business team, turning product requirements into shipped features. On pixis.ai I owned the infrastructure and releases.
The problem
A live platform with 23k+ commits of history, and the first job is knowing what you own. Access was spread across AWS, Kubernetes/ArgoCD and Firebase, and developers needed shared hosted infrastructure to run anything. Meanwhile the roadmap kept moving: UGC video, video scoring, credit billing and new models were all in progress.
What I built
The work ran in three layers: understand the system, make it safe to change, then ship through it.
- 01Understand
- Access audit: AWS, EKS/ArgoCD, Firebase
- Architecture and handover guides
- 02Make change safe
- One-command local stack
- Feature flags, staging first
- 03Ship
- ~170 PRs reviewed in 4 months
- Video, billing, model migrations
- 04Own the infrastructure
- pixis.ai on Pixis’s AWS, in Pulumi
- CMS on ECS Fargate and RDS
- 28 advisories cleared
Taking ownership
- Audited infrastructure and access across AWS, EKS/ArgoCD and Firebase, and wrote the architecture and handover guides.
- Built a one-command local environment for all three services: Dockerized Postgres, Redis, RabbitMQ and MongoDB, 7 async workers, and auto-detected webhook tunnels. A missing tunnel fails loudly instead of silently posting to staging.
Shipping safely
- Reviewed and merged ~170 PRs in 4 months: UGC video generation, video scoring, credit holds with atomic deduction, a video editor, and Kling 3 and Seedance model migrations.
- Added feature flags in backend and frontend, so new AI pipelines reached staging before production.
- Diagnosed a production bug where results collapsed from four images to one: Socket.IO events were lost across pods without sticky sessions. I shipped a self-healing client re-fetch.
pixis.ai and its CMS
- Moved www.pixis.ai from agency hosting to Pixis’s own AWS account on Amplify, in Pulumi, with GitHub Actions OIDC deploys and gated production.
- Re-platformed the Craft CMS onto ECS Fargate, RDS and S3/CloudFront in Pulumi, with migrations on deploy and a deep health check.
- Upgraded Craft and pinned patched PHP, clearing 3 critical and 25 high security advisories.
- Added pre-build schema validation and fail-fast builds, so a bad CMS deploy can’t take the site down.
Results
- Led 6 engineers and merged ~170 PRs in 4 months across 3 production repos.
- New AI features shipped behind flags, including the Pixis × Google co-development initiative deliverables, which I designed and my team built.
- pixis.ai runs on Pixis’s own AWS account, as code, with the CMS at zero known advisories.
What I’d do differently
- Flags and a local stack in week one. The local environment landed in May and feature flags in July, three months in. Both made every later change safer; I’d treat them as part of the takeover itself.
- Guardrails before the first incident. A schema mismatch once published an empty site, and the old health check stayed green while the API failed. The guardrails came after. Now I’d add them during the migration.
- Fix the root cause in the same cycle. The Socket.IO fix was a client-side stopgap pending a load-balancer change. I’d schedule both together, so the stopgap doesn’t quietly become the fix.