Our Work

Engagements That Held Up Under Production Load

Anonymized case studies from real production work — reliability, AI engineering, automation and platform.

8Anonymized Cases
4Domains Covered
100%Uptime Mindset
CloudflareWAFVarnishIncident Response
Read Case→

Edge DDoS Mitigation Under Election-Cycle Load

High-traffic political opinion publisher

Coordinated L7 attacks during a presidential election cycle threatened to take a national opinion publisher offline at peak ad-revenue hours. Designed and rolled out a layered edge defense without breaking caching for legitimate readers.

99.9%
Uptime during attack windows
<50ms
TTFB held at edge
0
Cache poisoning incidents
FinOpsCloudflareCache strategyCost optimization
Read Case→

Cutting CDN Spend Without Breaking the Reader Experience

High-traffic political opinion publisher

CDN bill was growing faster than revenue. Audited the cache strategy end-to-end, killed the long tail of pointless requests, and brought the bill back to a sensible ratio without touching reader latency.

~38%
Reduction in monthly CDN spend
0
Regression on TTFB or cache hit ratio
100%
Captured in dashboards, not anecdotes
PerformanceVarnishLighthouseCore Web Vitals
Read Case→

Lighthouse 100 on a Mass-Volume Regional News Portal

Regional news portal in Central-West Brazil

A high-volume regional news portal was hemorrhaging readers to slow page loads on mobile. Re-architected the cache layer, eliminated render-blocking, and pushed Lighthouse to a sustained green band — without rewriting the CMS.

32 → 95
Lighthouse Performance (mobile)
-1.7s
Time to Interactive
+18%
Sessions per user
WordPressDRAutomationMulti-tenant
Read Case→

Multi-Tenant WordPress Restore Pipeline

Internal SaaS operation

Manual WordPress restores were taking 6-8 hours per site and going wrong in subtle ways no one noticed until weeks later. Built a deterministic restore pipeline that completes in under an hour and verifies itself.

<1h
Average restore time (was 6-8h)
100%
Reproducible across sites
0
Silent restores after rollout
TelegramPlugin architectureBullMQMulti-tenant
Read Case→

41-Plugin Telegram Bot Platform with SaaS Multi-Tenancy

Tyber.io internal product

Built a Telegram bot platform that runs 41 independent plugins on a single grammY runtime, with per-tenant credential vaults and a plugin-as-product model. The bot itself is the deployment, distribution and billing surface for AI-powered ops apps.

41
Independent plugins in production
905+
Automated tests on the bot core
Multi-tenant
Per-user credential vault, no shared keys
ClaudeCode reviewAI agentsDevOps
Read Case→

AI Pull-Request Reviewer with Re-Review and Audit

Tyber.io internal product (productized)

Pull requests don't get reviewed if humans are busy, and humans are always busy. Built an AI reviewer that runs on every push, escalates only when needed, re-reviews on new commits and audits its own past comments — with a daily budget cap to keep the cost honest.

Per-push
Auto re-review when head_sha changes
Daily-cap
Hard $ budget on Claude API spend
Self-audit
Resolves stale comments after merge or rebase
Cloudflare WorkersEdgeMulti-siteR2
Read Case→

Edge DNS Router for Wildcard SaaS Sites

Tyber.io internal product

Built a Cloudflare Worker that routes `*.tyber.io` traffic to per-tenant content stored in R2 — including multi-page sites, auth-protected sites and reserved subdomains — without spinning up a backend for any of it.

<10ms
Worker latency at edge p95
Wildcard
Single Worker serves all subdomains
R2-backed
Zero origin pulls, no traditional server
DiscordDeploy automationMulti-tenantGuardrails
Read Case→

Multi-Team Discord Deploy Bot with Human-in-the-Loop Guardrails

Internal multi-team operation across regions

Five regional teams sharing one platform, each with its own deploy cadence, environment and SSH user. Built a Discord bot that lets each team deploy from chat without touching another team's infra — and refuses to do anything destructive without a human confirmation.

5
Isolated teams, single bot instance
Per-team
SSH user matches team, never `deploy`
Human-in-loop
Destructive ops require explicit confirm