Systems engineering for software that has to hold up under real traffic.

Rayder Systems is a software systems consulting practice. Runtime and performance, distributed systems and async Rust, production reliability and incident forensics, and the agent and evaluation infrastructure behind AI tooling. Every change is verified against a reference, so a faster or simpler system provably behaves the same.

hello@rayder.systems

Practice

Performance & runtime: Profile the hot paths, move the right ones to native code, fix memory layout and allocation behavior, and bring down p95 latency and request volume.
Distributed systems & async Rust: Peer-to-peer sync, request/response protocols, timeouts and backpressure, per-peer retry and feature detection, and the transport observability to see what is actually happening.
Production reliability & forensics: Diagnose deadlocks, async stalls, and runaway memory under real load with thread stacks, futex waits, and dump hooks. Then ship the fix, the watchdog, and a test that reproduces the failure.
Runtime, VM & compiler: Custom VM and C-runtime integration, compiler and interpreter parity, single-allocation deserialization, memory-mapped state, and native fast paths that match the reference exactly.
AI agent & evaluation infrastructure: Long-running agent harnesses, supervisors and heartbeat checks, append-only event buses, multi-model fan-out, and grading by reference and differential checks.
Technical due diligence: Code, architecture, dependencies, delivery risk, security posture, and execution quality, returned as clear written tradeoffs.

Selected work

Zorpproduction blockchain node · Rust

Peer-to-peer sync, runtime state migration, and async reliability on a production node. Every change verified against a reference; migration output hashed bit-for-bit identical.

p95 round-trip −60%

before164 ms

after66 ms

state migration −76%

before385 s

after91 s

sync requests −61%

before105,101

after40,555

Async stalls traced to overloaded responders and fixed with bounded queues and watchdogs, plus a stress harness that reproduces the failure.
TlonVere · Urbit's C runtime
- Built the khan and conn IPC drivers and the noun wire protocol across the kernel/runtime line; batched event-log replay on the boot path; wrote Nock jets matched exactly to the interpreter.
VaporwareUrbit · Hoon backend
- Token-gated software distribution: a chain-log scanner, an ERC-1155 and account-event parser, an ownership store, and an update stream for the apps on top.
EarlierML · data systems
- Fraud-detection ML in production at Feedzai; Elixir/Erlang healthcare data at Cota; Scala/Spark data platforms at Owl Analytics, Bank of America, and Morgan Stanley.

Approach

Direct technical diagnosis: what is actually wrong, not what is easiest to say.
Clear written tradeoffs you can act on.
Changes verified against a reference: differential runs, deterministic A/B, bit-identical state hashes where it matters.
Focused delivery on agreed scope.
Handoff notes that make the next operator's job easier.

Contact

Email hello@rayder.systems with a short description of the system and the problem in front of you.