blog928 posts
Engineering the millisecond cloud.
Deep-dives and how-tos on microVMs, Firecracker, snapshot-restore, and the infrastructure behind safe, fast AI agent execution.
928 posts
Top 10 Ephemeral Development Environment Platforms in 2026: What Delete Leaves Behind
Every platform here has a delete button and every one means something different by it. Ten ephemeral development environment platforms graded on the end of the lifecycle: the six classes of residue a teardown can leave, which ones each platform handles, and where the orphans collect.
Ajay KumarTop 13 Throwaway psql Platforms for Preview Environments (2026)
A preview environment is an app and a database that have to be born together, change together on every push, and die together when the pull request closes. Thirteen options graded on the four lifecycle events, not on how fast they hand you a connection string.
Ajay KumarWhat Is Vibe Coding? (And Where the Generated Code Should Actually Run)
Vibe coding is prompting a model for code, not reading it, and steering by whether the thing works. That's fine for a weekend toy. It is less fine when the unread code runs npm install on the laptop that holds your SSH keys.
Ajay KumarHow to Create an AI Agent Sandbox (Step by Step, 2026)
Isolation boundary, disposable filesystem, network policy, three timeouts, guaranteed teardown, and one run_code tool the model calls. Three ways to build it, with code for each.
Ajay KumarSlopsquatting: When Your AI Agent Installs a Package That Doesn't Exist (Yet)
Typosquatting waits for a human to mistype a real package. Slopsquatting waits for a model to confidently invent a fake one, and for the next model to invent the same name. Here is how to gate an agent's installs before the guess becomes your dependency.
Ajay KumarChecking Untrusted Proofs (Lean, Coq/Rocq, Isabelle) in Isolated microVMs
Proof checking looks like the safest workload there is, right up until you remember that Lean elaborates a file by running it. An RL pipeline generating millions of candidate proofs has two problems: what a proof can do to your machines, and what it can make your checker say.
Ajay KumarEphemeral microVM Environments for Hackathons and Coding Workshops
Two hundred attendees, forty broken Python installs, wifi that dies under npm install, one crypto miner "for the demo", and thirty projects to judge at 4pm. The event-day architecture: a machine per team, frozen at the bell.
Ajay KumarLLM-as-Judge vs Execution-Based Evals: Why Your Code Evals Need a Sandbox
A judge model will call wrong code idiomatic. Running it tells the truth, if the code can't edit the tests, exit early, or patch the assertion. Execute first, judge the remainder.
Ajay KumarDaytona vs Vercel Sandbox (2026): Which One for AI Code Execution?
Daytona grew up as a dev-environment platform; Vercel Sandbox grew up next to a deploy platform. Both now run code your model wrote. The question is whether your agent needs a workspace or a box shaped like a request.
Ajay KumarKVM Memslots Explained: How a microVM's RAM Is Actually Mapped
A guest's "physical" RAM is a range of ordinary host virtual memory plus one lookup table. That table, the memslot list, is why snapshot restore, copy-on-write and memory streaming are cheap.
Ajay KumarTop 14 Ephemeral Development Environment Platforms in 2026: The Resume Path Nobody Benchmarks
Every platform in this category publishes a create number. None of them publish the one you hit ten times a day: the wait after the platform decided nobody was looking. Fourteen platforms graded on what idle actually means -- stop, suspend, hibernate or destroy -- what survives it, what it costs at rest, and whether coming back is the same code path as create or a slower one nobody measured.
Ajay KumarTop 12 Disposable psql Platforms for Staging in 2026: Graded on Fidelity to Production
CI wants a database that is fast and empty. Staging wants one that behaves like production when a real app is pointed at it for days, and then goes away by itself. Twelve options graded on superuser and config control, extensions, 200 idle connections overnight, copies at an instant, long DDL and replication slots, and what idle actually costs.
Ajay KumarRunning a DICOM Pipeline in a MicroVM: Untrusted Medical Images, One Study at a Time
A DICOM file tells your process which several thousand lines of C to run next, and hospital archives are full of objects written by a decade of vendor encoders — so malformed is the normal case, not the attack case. Here is the per-study microVM shape that makes a crashing codec one deleted machine instead of a disclosure question.
Ajay KumarPer-Tenant Map Tile Rendering in MicroVMs: When Your Customers Write the Stylesheet
A map style is not configuration, it is a program: a few kilobytes of JSON that tells thirty years of C++ how much memory to allocate and which files to open. Here is what happens when one tenant writes a bad one, and the per-tenant microVM shape that keeps it their problem instead of everybody's.
Ajay KumarPer-Tenant Payroll and Tax Engines in microVMs: Determinism Is the Product
A payroll or tax engine is a pure function on paper and a function of fifteen undeclared arguments in production: tzdata, ICU, glibc, a rounding mode nobody chose, and the customer's own deduction formula. Then an amended return arrives three years later and asks for the same number, plus an explanation.
Ajay KumarKVM Halt Polling Explained: Why an Idle microVM Can Burn a Core Doing Nothing
A guest with nothing to do executes HLT, and the host has to decide whether to block the vCPU thread or spin for a moment hoping an interrupt arrives. That decision is a module parameter, it is global to the host, and on a fleet of sleeping microVMs it quietly puts a floor under every CPU number you bill on.
Ajay KumarEROFS vs SquashFS for a Read-Only microVM Rootfs: Compression You Pay For on Every Page Fault
A microVM rootfs is almost entirely read-only, so compressing it looks like free money. It is not free — it is relocated. Compression moves the cost of those bytes off the network and onto the CPU of whichever guest faults the page in, and whether that is a brilliant trade or a self-inflicted wound depends on which resource you were actually short of.
Ajay Kumarvirtio-net Offloads and Why Your microVM's Network Is Slower Than the Wire
Bulk transfers look fine. A sixty-four-byte request/response benchmark gets a third of bare metal. Both numbers are correct, and the gap between them is a virtio feature negotiation, a TAP device with its own opinion, and a per-packet tax that every sandbox on the host pays at once.
Ajay KumarReadahead Inside a microVM: The Kernel Guessing Wrong, 4 KiB at a Time
There is no disk. A guest read is a descriptor on a virtqueue, an exit to the VMM, a file read on the host, and a second page cache doing its own speculation underneath yours. Both layers are guessing about the same access pattern, and only one of them is paying for the exits.
Ajay KumarFlamegraphs Inside a microVM: When perf Has No Counters to Count
The flamegraph was 1,400 pixels wide and ninety-six percent of it was one box labelled [unknown]. perf runs fine inside a microVM; the recipe everybody memorised depends on three things a guest quietly does not provide, and perf will not tell you which one bit you.
Ajay KumarTop 13 Ephemeral Development Environment Platforms in 2026: Who Is Allowed to Call Create
This category was designed for a person opening a laptop. In 2026 the fastest-growing caller is a webhook, a CI matrix or an agent. Thirteen platforms graded on whether create is a real API, what the isolation boundary is when a model wrote the code, what happens at fifty concurrent, and whether there is a cold path at all or only a warm path you can miss.
Ajay KumarTop 11 Ephemeral Postgres Platforms for CI Pipelines: Time to First Connection, and Whether It Is Real Postgres
Your CI job does not care about QPS. It cares about two things nobody benchmarks: how long until a connection is actually accepted, and whether the thing accepting it is real Postgres or something wire-compatible. Eleven options graded on those, plus isolation, leak cost and whether a seeded dataset can be reused without re-seeding.
Ajay KumarA Font Is a Program: Rendering User-Uploaded Typefaces Without Trusting Them
The moment you accept a .ttf from a user, you have accepted a program: TrueType carries a stack machine, CFF carries another one, and woff2 puts a Brotli decoder in front of both. Here is why that parser should not run in the process holding your database credentials, and the microVM shape that fixes it.
Ajay KumarCompiling a Stranger's Shader: Isolate the Toolchain, Because You Cannot Isolate a GPU You Do Not Have
If your product accepts shader source from users or from a model, you are running a pile of C++ parsers on hostile input. PandaStack has no GPU and cannot run your shader on one — but glslang, spirv-val, spirv-opt and spirv-cross are all CPU work, and that is the part that actually hurts you.
Ajay Kumar