NAYEEM
FARDIN

It means I care about the systems underneath the AI, not just the prompt on top of it. Wrapping a model in an app is the easy part. The engineering is in the queues that keep inference off the request path, the retrieval that has to return the right passage, the boundaries that stop a model touching things it should not, the metrics that tell you it is degrading, and the cost of every GPU second. In DeliveryLens the model writes the prose but never carries a number, because a small model once mis-transcribed a year out of a correct result. That rule is what the title means to me.

Because a demo gets to assume the model is fast, available, and right. Real systems do not. The interesting questions all start where the demo ends: what does the user see while inference is running, what happens when the provider rate-limits you, how do you tell a wrong answer from a slow one, and what does this cost at a hundred times the traffic. I gravitated to backend and infrastructure because that is where those answers live. AI made the questions more interesting, not different.

Four habits, mostly learned by getting them wrong first. Keep the model off the critical path — in RetailOS Lite a visit submission enqueues work and returns, so a slow model never blocks a field rep. Give the model narrow boundaries — YOLO produces grounded detections, an LLM reasons about them, and typed tools running SQL produce every number, so nothing measured depends on the model remembering it. Make degradation loud — when the Cohere reranker times out, retrieval falls back to a diversity pick and records which path ran, so a weaker answer is flagged rather than shipped silently. Spend cheap checks before expensive ones — the fraud engine runs five cheap signals before anything reaches a GPU.

Ideally I can answer "what broke, when, and for whom" without guessing. That means queues with dead letters and a replay CLI that defaults to a dry run, correlation IDs threaded across services, Prometheus and Grafana with Loki for logs, and results that carry provenance — when each source last synced, what was excluded, which tasks could not be judged at all. My homelab is where I practise the unglamorous half: 19 containers with an incident log and real postmortems, and a weekly job that restores a random sample of the backup and checksums it, because a backup you have never restored is a guess.

Because that is where things actually break, and I find that part genuinely interesting rather than a chore. Spot instances need graceful eviction or you lose work. A 65-branch repository will exhaust GitHub's hourly rate limit unless the sync is incremental and watermarked. Container images pinned by tag can silently change what runs, so mine are pinned by @sha256: digest. A shingled drive stalls for seconds under random writes, so write-heavy data belongs on the SSD and only sequential bulk goes on the HDD. None of that shows up in a feature list, and all of it decides whether the thing stays up.

RetailOS Lite — an AI-native retail execution platform with async shelf analysis, a five-signal fraud engine that runs before the GPU, and a RAG assistant that queries exact data before it reaches for vectors. DeliveryLens — a delivery-intelligence assistant where 16 typed tools run SQL so no measured number is ever written by a model. Deskemy — the one with real users. ShelfWatch — dense retail detection on AWS EKS, 4× faster CPU inference and 70–90% cheaper through INT8 quantization and Spot diversification. PixTools — a Go HTTP edge over Celery workers on a Terraform-provisioned K3s cluster, with circuit breakers and idempotency keys for when things go wrong.

Deskemy, an offline course player for Windows, because I have now solved the same problem in it three times. It plays through libmpv rather than an HTML5 video element, so it handles whatever codecs a downloaded course ships with — but then something has to draw the controls over that video. The first version put the video in a separate window above a web view, and the picture lagged the layout on every resize. Version 1.x composited the video beneath a transparent web view, which fixed it but only on Windows. 2.0 rebuilt the whole interface in Slint so the video renders into the UI's own OpenGL context and the problem simply stops existing. It has 52 stars, 343 downloads, and a community-maintained Chinese translation, which I did not expect.

Operating things properly, rather than only building them. Most of that happens in a two-machine homelab running 19 rootless podman containers as systemd Quadlets, which has taught me more about SELinux labelling, split-horizon DNS, mount ordering, and digest-pinned images than any tutorial would have. I am working toward platform and reliability engineering — deeper Kubernetes than I have needed so far, and the distributed-systems fundamentals behind consensus, backpressure, and failure detection. I would rather say that plainly than pretend I have finished.

AI infrastructure and platform engineering — the layer that makes model-backed features dependable rather than impressive once. Inference serving and the queues around it, retrieval systems that have to be right and not just plausible, and the observability that tells you which of the two you have. More broadly: backend and distributed systems where correctness under failure is the actual requirement, and developer platforms where one person's work removes friction for everyone else. Cost-aware infrastructure interests me too, because a system nobody can afford to run is not finished.

I graduate in September 2026 and I am looking for engineering roles in AI infrastructure, backend, or platform work — new-grad positions and internships both. I am also happy to talk about project collaborations, or about any of the systems on this site if something here is useful to you. The fastest route is email at nayeemfardin00@gmail.com; my code is on github.com/NFRohan and I am on LinkedIn.