COMPANY PROJECT · FILUM AI · 2024 — 2026
Filum CX
An omnichannel customer engagement platform: one inbox for 10+ channels, AI agents and automations, and support workflows for customer service teams.
- Next.js
- Node.js
- Python
- GCP Cloud Functions
- Cloud Run
- Redis
- TanStack Query/Table
- Ant Design
Built at Filum AI as part of the engineering team. The product belongs to the company; you need a company account to sign in.

At a glance
Filum CX is Filum AI’s main product. Businesses use it to talk to their customers on every channel from one inbox, with AI agents and automations handling the repetitive work.
I worked on it for two years in a 15-person engineering team. I designed and led the reliability layer for the integrations, owned the automation work, and added features to conversations and tickets. This page covers the problems I took on and what changed because of them.
Making channel integrations reliable
Filum CX supports more than 80 integrations, running on Cloud Functions. When a channel integration broke, the customer was the first to notice. They told our customer service team, CS checked and passed it to the developers, the developers traced the problem, and then the answer travelled back to the customer.
I designed the system that changed this for more than 10 channels, and built most of it.
| Kind of failure | Before | After |
|---|---|---|
| Expired credential (token, login) | Found by the customer. 2–3 hours of back-and-forth for an experienced developer, about half a day for a newer one, then the customer re-authorizes | The customer sees the connection is about to expire, or has expired, and re-authorizes it themselves, before anything stops working |
| Infrastructure (Cloud Function out of memory, traffic spike, channel API errors) | Half a day to a full day to find the cause, then the missed data had to be re-run by hand for the customer | The cause is visible right away in centralized logs. Failed jobs retry automatically, and if a retry still fails, the team gets an alert in its internal channel |
- Credential monitoringTracks the state of every connection and shows customers when one is close to expiring, so they can re-authorize it in time.
- Centralized loggingLogs from all integrations in one place, so a failure points straight at its cause instead of a long trace.
- Automatic retriesData that failed to sync is processed again without anyone re-running it.
- Internal alertsWhen retries are not enough, the team is told first, often before the customer would notice.
The hardest bug: a view count that was always wrong
The conversation view count, the number of conversations in each view of the inbox, had been wrong for as long as anyone could remember. It got fixed again and again, and a few days later a customer would report it was off again. Nobody had found the root cause.
I traced it to how the count was built: it was calculated on the frontend, with no source of truth from the backend, so it drifted as conversations changed. Patching the frontend could never hold.
I added a flow where the backend sends a snapshot of the counts to the frontend every time something changes. Customers stopped reporting wrong counts after that.
Automation and AI
AI replies to Facebook comments
Set up an automation where AI answers customers’ Facebook comments. Before, someone had to watch the page and reply by hand. Now no comment goes unanswered.
An Agent node for any workflow
Built an Agent node that can be invoked inside other automation flows, so AI agents become a building block customers can drop into their own workflows.
Triggers and actions
Helped build the automation system with 10+ triggers and 10+ actions.
Conversations, tickets, and speed
SLA and attributes
Added SLAs and custom attributes to conversations and tickets, so support teams see what is about to breach and can classify requests.
15 s to about 8 s
Reordered the API calls: what loads first, what can wait, and what is not called until the user opens that part. Page load for customers dropped from 15 seconds to just over 8.
Lighter frontend
Lazy loading, tree shaking, list virtualization, and render optimization to cut wasted work in the browser.
10 channels I integrated
Of the 80+ integrations the platform supports, I integrated 10 channels myself, including Messenger, Zalo OA, and 3CX, into the central conversation manager.
How it works
- FrontendNext.js with Ant Design, shadcn/ui, and TanStack Query/Table.
- IntegrationsNode.js and Python services on GCP Cloud Functions and Cloud Run, with Redis.
- ReliabilityCentralized logging, credential monitoring, retries, and internal alerts for 10+ channel integrations.
- SecurityAI-assisted security audits, including finding GCP access tokens exposed through custom tool APIs. Read the write-up →
See it live
Opens the live site in a new tab. You need a company account to sign in.