Mubdiur Rahman — AI Platform & Reliability Engineer — designs platforms that learn production portals, run near-realtime checks, and alert on what matters. AI is the force multiplier; reliability is the job; safety is non-negotiable.
The monitoring industry runs on humans staring at dashboards for eight-hour shifts. This operator’s answer is automation: platforms that learn a system’s structure once, then run checks back-to-back with no polling gap — catching failures before they ever reach an error log. The tooling is browser-grade; the discipline is production-grade.
The same discipline runs the infrastructure underneath: Docker and Traefik hosts with Let’s Encrypt TLS, zero-downtime rebuilds, and server backends guarded by rate limits and SSRF checks. Guardrails — injection defense, PII redaction, audit trails — sit between every model call and the data. MCP tooling and AI-assisted workflows are used where they earn their keep: as force multipliers, not demos.
Browser automation at scale — Playwright, Cypress, Appium. Shared browser instances with strict lifecycle management, deterministic resource use, and structured extraction from live UIs.
Surveillance of Dynatrace problem boards, Kibana error views, Datadog cost panels, and transaction portals. eBPF probes and OTel spans under the hood — near-realtime, back-to-back checks with append-only results.
Docker and Traefik hosts with Let’s Encrypt TLS and zero-downtime rebuilds. OpenStack and bare Linux underneath, server backends guarded by rate limits and SSRF checks. Hardened to least privilege.
Guardrails in the loop: prompt-injection defense, PII redaction, output validation, allow-listed tools. Agents get capabilities, not keys — and every action is audited.
Agents as state machines with explicit failure paths — not prompt spaghetti. Versioned prompts, typed tool contracts, and architecture that survives the engineer who built it.
MCP tooling in production stacks, agent-driven automation, and eval-driven iteration. AI is a force multiplier and an on-call copilot — never a demo.
Node.js and TypeScript services with zero unnecessary dependencies. Git and CI/CD workflows, terminal-first operations, rapid prototyping with production discipline.
crawl4ai-powered intake and Playwright-driven extraction — map a portal once, pull structured records back-to-back. Built for hostile UIs and shifting layouts.
Every prompt, model, and tool change ships behind a regression suite. Shadow traffic, canary deploys, automated rollback, drift detection — the AI equivalent of code review.
This newspaper is not a template — it’s a production system. Next.js 16 standalone behind Traefik with Let’s Encrypt, a Docker-based rebuild-and-deploy pipeline, and a server-tool backend with rate limiting, timeouts, and SSRF guardrails. 80+ utilities and a live instrumented dashboard ship from the same repository.
SLO-based paging, burn-rate alerting, runbooks, and MTTR tracking — the discipline of being paged less while knowing more. Automation handles the noise; humans handle the judgment.
Long-running Node.js daemons that run unattended — messenger-based operations with self-confirming delivery, structured logging, and recovery built in. Set-and-forget systems that report back.
Next.js 16 standalone behind Traefik with Let’s Encrypt. Docker-based rebuild and deploy pipeline, plus a generic server-tool backend with rate limits and network guardrails.
A multi-project Docker host operating as the backbone for everything shipped. Containerized isolation, least-privilege access, Let’s Encrypt TLS across the board.
Copy time in mm/dd/yyyy hh:mm AM/PM across UTC, PT (PDT/PST), and ET (EDT/EST) — DST-aware, for now or any custom moment
Live TLS handshake + full X.509 dissection — validity to the millisecond, fingerprints, SANs, SCTs, chain, raw DER
Validate and format JSON with detailed error messages — line, column, and position
Generate QR codes from text or URLs
Engineering managers don’t fear AI. They fear the bill, the 3 a.m. page from something nobody understands, and the prompt that silently changed behavior in production. All three are engineering problems — not magic problems.
Cache the tokens, compress the prompt, gate every change behind a regression suite, shadow-test the next candidate against live traffic, and put guardrails between the model and your data. Safe, maintainable, affordable — in that order. That’s the whole platform.
PRODUCTION ......... PARTLY CLOUDY · 99.95% SUNNY DATACENTER ......... HOT. ALWAYS. TOKEN BUDGET ....... FALLING (−40% POST-CACHE) CHANCE OF INCIDENT . 0.05% — AT 02:00, PROBABLY ALERT FATIGUE ...... ZERO. AS IT SHOULD BE.
S R E L L O G V B X P Q T A K N O I R H W C A A V Q M Z U B F E P 9 5 G W N Y J C H T D S A X E G P U A L M V R O E I Y K B J C A L E R T R W H D L O K I T Q N M S U F R Z P X F C G V J K L O I T E S Y M C P B A D N F O W C H L Z U Q T J M T B H D X N E K A R Y S C M Z S L O W F J V G T
Platform, automation & reliability engineering opportunities. Will trade architecture walkthroughs and a full dossier for a conversation.