Remote C2C machine learning compiler and performance engineering contracts are sourced mostly through niche staffing vendors who hold direct MSP or VMS access to chip companies, hyperscalers, and inference startups — not through public job boards. You find them by tracking a short list of specialist vendors, monitoring GitHub/LLVM community signals for team expansions, and applying through your own C2C corp within hours of a requirement dropping, because these reqs rarely stay open long.
If you're a compiler or performance engineer who's spent years in LLVM, MLIR, XLA, or Triton internals, you already know the regular job boards are useless for this work. LinkedIn shows you thousands of generic "ML Engineer" postings and maybe two that mention kernel fusion or instruction scheduling. Indeed is worse. The actual demand for people who can make a matmul run faster on a new accelerator is real and growing, but it moves through a completely different pipeline: corp-to-corp staffing desks that specialize in silicon and infra talent.
This piece is for the engineer who's tired of scrolling irrelevant postings and wants a working method — not a pep talk.
Why ML compiler and performance engineering pays differently than regular ML contracting
Compiler and performance engineering sits at the intersection of two scarce skill pools: classic systems/compiler expertise (LLVM, MLIR, register allocation, autotuning) and modern ML workload knowledge (transformer inference, quantization, kernel fusion, memory bandwidth bottlenecks). Very few engineers have both. That scarcity is why rate sheets for this niche run meaningfully above generic "ML Engineer" or "Data Scientist" C2C postings — the buyer's alternative isn't a cheaper contractor, it's missing a chip launch deadline.
Because the pool is small, clients lean on trusted staffing relationships instead of open postings. A hiring manager at an accelerator startup would rather call three vendors who've placed compiler people before than post publicly and sort through hundreds of resumes that don't match. That's good news for you if you get inside that vendor loop, and bad news if you're relying on public boards alone.
In short: this market rewards engineers who go direct to specialist vendors and speak the domain language fluently in the first email, not those who mass-apply.
Where remote C2C ML compiler contracts actually get posted
Public job boards catch a small fraction of these requirements, and usually late. The real flow runs through four channels, in order of value:
- Specialist staffing vendors with MSP access. These are boutique firms, often five to thirty recruiters, that hold vendor management system (VMS) seats with chipmakers, hyperscaler infra teams, and well-funded inference startups. They see the requirement the moment it's submitted internally.
- Direct-to-hiring-manager outreach. Compiler teams are small and their leads are usually reachable. A cold message referencing a specific PR you shipped to LLVM or a Triton kernel you optimized gets read, because these managers know how few people can do this work.
- Community signal tracking. LLVM Discourse, MLIR mailing lists, and Triton's GitHub discussions often show a team expanding before any req exists publicly. Watching who's hiring reviewers or merging infra changes tells you who's about to need contractors.
- General C2C job boards, last. Dice, ClearanceJobs (for cleared work), and general VMS portals do carry some of these reqs, but usually after the vendor has already run its own bench and network first.
Plain summary: the best compiler contracts move through relationships and specialist vendors before they ever hit a searchable board, so your job is to get in front of the vendor, not the posting.
The vendor hotlist: who actually staffs this niche
You won't find a single definitive public list — vendor relationships shift and firms rebrand — but the pattern holds across the market. Look for staffing firms with three traits: they list "compiler," "kernel," or "performance engineering" as a named practice area (not buried under generic "AI/ML staffing"), they've placed people at chip or inference companies you can verify on LinkedIn, and their recruiters ask technical follow-up questions instead of reading a script.
A fast way to build your own hotlist:
- Search LinkedIn for "compiler engineer contract" and filter by people who list a staffing agency as their employer of record — their recruiter contacts are visible.
- Check who's posting compiler/performance roles on Dice and ClearanceJobs, then note the agency name even if the specific req isn't a fit. That agency likely runs more of the same.
- Ask other C2C compiler contractors directly. This is a small world; most people in it know two or three vendors worth working with and are willing to share names once you're clearly a peer, not a lead.
Once you have five to ten vendor contacts who genuinely staff this niche, treat them like a sales pipeline. Send updated availability every few weeks. Recruiters at specialist desks remember engineers who stay in touch without being pushy, and they'll call you first when the right req lands.
How C2C rates work for ML compiler and performance roles
Rate structures in this niche follow the standard C2C chain — end client pays a bill rate to the prime vendor, the prime marks it up and passes a rate to your corp, sometimes through a sub-vendor layer that adds another markup. The more layers between you and the end client, the thinner your take. This is exactly why understanding rate sheet discrepancies matters more here than in generic IT staffing — a vendor who won't disclose the bill rate on a compiler req is almost always hiding a heavier markup than the work justifies.
| Factor | Impact on your rate |
|---|---|
| Direct vendor (1 layer to end client) | Highest pass-through, less margin lost |
| Sub-vendor chain (2+ layers) | Meaningfully reduced take-home for same bill rate |
| Specialized runtime/hardware (custom silicon, novel accelerators) | Premium over general CUDA/LLVM work |
| Security clearance or on-prem hybrid requirement | Premium, but limits your applicant pool advantage as fewer contractors qualify |
| Short-term backfill vs multi-quarter engagement | Short backfills often pay a premium for urgency; long engagements trade some rate for stability |
Before agreeing to a rate, always ask the vendor for the bill rate range, not just what they're offering your corp. A vendor confident in a fair markup will tell you. One that deflects is a signal to negotiate harder or walk.
How to position yourself for compiler and performance C2C work
Generic resumes lose in this niche because the reviewer is often a technical lead skimming for specific signals, not a generalist recruiter. Your resume and outreach need to read like a compiler engineer wrote them, not a staffing template.
- Lead with the specific stack. Name the exact frameworks: MLIR, XLA, TVM, Triton, LLVM backends, CUDA, ROCm. Vague terms like "ML infrastructure" get filtered out by people scanning for keywords they actually use daily.
- Quantify the performance win, not the project. "Reduced kernel latency" says nothing. "Cut inference latency through operator fusion and memory layout changes" tells a compiler lead you understand the actual levers.
- List hardware targets explicitly. GPUs, TPUs, custom ASICs, FPGAs — say which. Teams building for novel silicon want engineers who've dealt with unusual memory hierarchies before, not just NVIDIA-only experience.
- Reference open-source contributions if you have them. A merged PR to LLVM, MLIR, or Triton is worth more than a paragraph of bullet points to a technical reviewer.
- Register your corp with a clean rate structure ready. Have your C2C entity, insurance, and standard rate range prepared before you're deep in a vendor conversation. Momentum dies waiting on paperwork.
- Apply within the first hours of a posting. Compiler and performance reqs in this niche close fast because the buyer's pool is small and vendors move their best-fit candidates immediately. Waiting a day can mean the req is already staffed off someone's bench.
Plain summary: your materials need to prove domain fluency in the first ten seconds of a scan, and your process needs to be fast enough that you're never the reason a good req slips away.
Speed matters more in this niche than in general C2C markets
Because the compiler/performance talent pool is small, vendors don't run open pipelines with fifty candidates per req the way generic staffing does. They usually have two or three people in mind before the req is even fully written, drawn from their bench and recent conversations. If you're not already warm with the vendor when the req appears, you're competing for the leftover slot, if there is one.
This is where being first with the right vendor beats being the most qualified candidate who applies late. It's also why manually monitoring five or six boards plus vendor inboxes is a losing strategy for a working contractor who's also delivering on a current engagement. An automated layer that watches for fresh compiler and performance postings the moment they go live, and gets your application or outreach in ahead of the pack, closes that gap. That's the entire premise behind GiraffyReach — detecting postings early and acting before hundreds of others even see them, in a market where "before" is measured in hours, not days.
If you want to understand how compiler engineering differs from adjacent performance roles before you position yourself for either, read ML Compiler Engineer vs Performance Engineer: What's the Real Difference? first — it clarifies which title to lead with when a vendor asks what you actually do.
Where GiraffyReach fits into your search
None of this replaces relationships with good vendors or a resume that speaks the domain's language. What it does is remove the part of the job where you're refreshing boards at midnight hoping to be first. GiraffyReach tracks fresh postings across the C2C market, including the specialist channels compiler and performance work flows through, and applies or reaches out before the queue fills up. For a niche where the difference between landing a contract and missing it is often a matter of hours, that head start is the whole game.
Be first, or be forgotten — nowhere is that more literal than in a market this small.