Cloud & platform engineer · Calgary, Alberta
I keep things running that are not allowed to fail — and I prove the checks can fail before I trust them. Five years in cloud infrastructure and telecom, including 2.5 years running 5G core networks at Nokia against a 99.9% uptime guarantee.
Considered for Cloud Engineer · DevOps Engineer · Platform Engineer · Site Reliability Engineer · Infrastructure Engineer · AI / ML Platform Engineer
github.com/sadvi11 LinkedIn sadhvisharma763@gmail.com Download CV (PDF)
A check that cannot fail is worse than no check.
It produces confidence without coverage. So every guard I ship gets removed on purpose in CI, and the build fails if the tests still pass without it. A green badge means nothing until you have watched it go red for the right reason.
An MCP server any AI assistant can call. It reads a Terraform plan and refuses to bless a dangerous one — the verdict is a Python comparison, so no wording in the plan argues it down.
There is a test where the resource is named
definitely_safe_to_delete_approved_by_cto and it still blocks,
because the name was never an input to the comparison. Then I attacked my
own tool and found that the attacker-controlled resource name was printed
raw into the text the model reads — so they could not change the verdict,
but they could forge a fake system instruction around it. Fixed at parse
time, which covers all five tools rather than one.
Twelve workloads — eight microservices in five languages, two databases, a cache and a message broker. Terraform builds the cluster; one templated Helm topology renders all twelve rather than twenty-four copy-pasted manifests.
The databases use Recreate, not RollingUpdate,
because a ReadWriteOnce EBS volume attaches to one node and a rolling
update deadlocks. Accepted downtime, chosen deliberately rather than
discovered at 2am. It took nine CI runs to go green and the commit history
says why — including that a container running as uid 0 with all
capabilities dropped is not root.
An EC2 inventory reporter, and the same reads behind an LLM tool-use loop.
The read-only contract is enforced by a Python allowlist on botocore's
request hook — not by a sentence in the system prompt. A model talked
into calling TerminateInstances still cannot: the call never
leaves the process.
Almost every Canadian posting I read names a compliance regime — FedRAMP and NIST in defence, HIPAA and ISO in health, OSFI in finance. The common problem underneath them is the same: prove the control exists, and prove it is enforced rather than documented.
Data residency is the version of that problem I have solved. It is a legal question before it is an architectural one, and most reference architectures are written for somewhere else — so I enforce it in policy rather than in a convention people are asked to remember.
Sectors this maps to: financial services · healthcare · government and public sector · energy. Control mappings written for SOC 2 and PCI DSS as well, in nokia-5g-to-aws-migration. aws-multi-account-governance · iac-security-guardrails
Every one of these was caught by me, in my own code, and fixed with a test that goes red if the fix is removed. None of it reached a user. They are here because finding your own faults is harder to fake than a green badge — and it is the difference between an engineer who is careful and one who only looks careful.
A policy that matched nothing, while the suite stayed green.
Terraform emits "tags": null, and not r.values.tags is false for null. I fixed the rule, then restructured the tests so one dead policy cannot hide behind its neighbours.
A green CI badge over a red test suite.
continue-on-error: true, with a comment claiming credentials were needed. Eleven of twelve passed offline.
A deployed model scoring 2/6 on real inputs at ~0.52 confidence.
Cross-validation had said 0.471 — worse than random. It now abstains when the input is outside its vocabulary instead of guessing.
A Lambda parsing Kinesis records on an SQS pipeline.
It also swallowed failures and returned success — which, on an SQS trigger, deletes the message instead of retrying it. Nothing tested that handler. The untested function was the broken one.
A prompt-injection hole in my own MCP server, two hours after publishing it.
The first test I wrote for it asserted the wrong property — that the hostile words disappeared. They belong to a real resource name; deleting them would misreport the plan. Only the forged structure had to go.
Most cloud engineers learn high availability from a course. I learned it on call, on a telecom core network, where the failure mode is not a red build — it is someone's call dropping mid-sentence.
Dec 2022 – Jul 2025 · Nokia, Delhi
5G core network functions — AMF, SMF, UPF, NRF — as containerised workloads on Kubernetes and OpenStack, in carrier production against a 99.9% SLA, for Bell Canada, T-Mobile US and Orange. Delivered a wave-based multi-site 4G→5G migration with zero subscriber downtime. Production on-call with structured root cause analysis.
Oct 2025 – present · Meta, via Influence Marketing · Calgary
Technical enablement on Meta AI products and XR hardware. Built and delivered training for 50+ retail staff. The infrastructure work above is built outside working hours.
Education & certification
Full merit scholarship. B.Tech, Shri Mata Vaishno Devi University. Microsoft Azure Fundamentals (AZ-900). AWS Solutions Architect Associate — studying, sitting September 2026.
The hardest part of an incident is rarely the fix. It is telling people what is happening while it is still happening — and writing it up afterwards so the next person is not starting from nothing.
Toastmasters International · District 42
Accountable for four to six clubs — their membership, their health, and their results — with no authority over any of them. Twice-yearly visits to each club with written reports, Area Council meetings, and running the area speech contests. Toastmasters describes the role as supportive rather than directive, which is a polite way of saying you get results by persuasion or you do not get them.
The contest runs club, then area, then division. Reaching second at division means winning at club level, winning the area, and then placing against the champions of the other areas in the division — a field drawn from more than a dozen clubs. Judged on content, structure and delivery, from memory, to time, in front of a room.