Four tracks, four contracts. Each is scoped, fixed, and shipped with the artifacts you need to maintain it after we leave, no retainer required, no black boxes. Read the page, then open a ticket.
SYNOPSIS. A work-stealing, quota-aware scheduler you can reason about line by line. Per-tenant fairness, deadline classes, and a replay log that lets you rerun any incident exactly as it happened.
DESCRIPTION. We've shipped four in production, two trading desks, one hardware test fleet, one genomics pipeline. The core is a few thousand lines of Rust with a single lock-free run queue per core. Everything that affects scheduling order is logged, so "why did this task run late?" is a query, not a guess. Tenants get hard quotas and soft bursts; starvation is provably bounded.
SYNOPSIS. Small, embeddable storage engines for products that cannot ship a Postgres dependency. Append-only log, snapshot + compaction, and a replication transport written from scratch.
DESCRIPTION. CRDT-friendly merge semantics, crash-safe by construction, and tested the way storage should be: a fuzz harness against a reference model plus property tests for every invariant. Ships as a single Rust crate with no runtime allocator surprises. Third-party audit available at your option, we'll book it and hand you the report.
SYNOPSIS. Custom wire formats, RPC dialects, and embedded-device protocols, specified first, in ABNF, with conformance test vectors before a byte of implementation ships.
DESCRIPTION. A spec you can hand to a second vendor and get an interoperable implementation. We design for forward compatibility, explicit framing, and machine-checkable grammars. We also do the unglamorous job: reverse-engineering an undocumented legacy protocol from disassembly so you can keep talking to a 2007 controller that nobody dares replace.
SYNOPSIS. Two weeks. Pick the path that's costing you money. We profile it, rewrite it, and prove the win with reproducible benchmarks in front of you.
DESCRIPTION. No retainer, no new dependency, no architecture astronomy. We instrument with perf and eBPF, find where the time actually goes, and fix the hot path in your codebase. You get a written report with before/after numbers and a script that anyone on your team can re-run to confirm them. If we can't beat your baseline, you don't pay the success fee.