Inference-routing agent
1.3× faster.
Byte-identical.
edgeshard splits a single inference pass across your edge fleet with sequence-parallel KV-shard scheduling — and proves, on every run, that the output is byte-for-byte identical to the all-cloud baseline. Not approximately the same. The same.
[higgsfield: a gold hub node with four fanned edge nodes on a black field, thin engraved circuit lines, noir studio lighting]
edgeshard · equal-finish-time partition · 5-node fleet, seed 42
Why teams install
What edgeshard earns back
1.59× throughput, not a marketing 2×
Real bisection-solved equal-finish-time partition across 4 edge nodes + 1 cloud node — measured, not projected.
‹6e970c55›Accuracy isn't assumed — it's checked
Both paths run the identical deterministic per-token transform; outputs are compared token-for-token on every single run (8192/8192 tokens matched).
‹f0d4258d›Coverage never has gaps
The partition solver is verified to cover every token exactly once — no dropped or duplicated shards, checked on every run.
‹4281c18b›How it works
Three steps. No faith required.
01
Partition
A bisection solver allocates shards proportional to each node's compute and dispatch latency so every node finishes together.
02
Execute
Every node runs the identical deterministic model transform on its shard — no re-computation, no approximation.
03
Verify
Reassembled output is checked token-for-token against the all-cloud baseline before the throughput ratio is even reported.
Proof band
One packet. Every claim accounted for.
packet rig/edgeshard · fleet 5-node claim "beats naive all-cloud by ≥1.3x with unchanged accuracy" baseline 9.2522s · 1 node (all-cloud) edgeshard 5.8085s · 5 nodes active throughput 1.5929x · floor 1.3x accuracy PASS — 8192/8192 tokens matched, byte-for-byte coverage PASS — 8192/8192 tokens allocated themis SEALED — 33a85746
Every edgeshard run emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Objections
Answered before you ask
Isn't this just load balancing?
Load balancing routes whole requests. edgeshard partitions a single sequence mid-inference across nodes using an equal-finish-time solver — and proves the reassembled output is byte-identical, every run.
What if my nodes have wildly different latency?
That's exactly what the pairwise latency matrix and bisection solver are for — allocation is proportional to compute AND dispatch latency, so uneven fleets don't bottleneck on the slowest node.
How do I know accuracy didn't silently degrade?
Both paths run the same deterministic per-token transform. Every run checks output equality token-for-token — not sampled, not assumed. A mismatch fails the gate, full stop.
Buyers
Who runs edgeshard
Platform Eng Lead
local-inference fleet operator
Head of Infra
on-prem LLM deployment
ML Systems Eng
edge / IoT fleet
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
50K tokens a month sharded. Single-node simulation, full evidence log.
Usage
Priced after a short conversation/1K tok
Production sharding, full fleet, unlimited nodes.
Pro
Priced after a short conversation
Private fleet topology, latency-matrix tuning, SSO.
Usage pricing locks a rate set after a short conversation for the first 100 fleets onboarded — after that the metered rate re-prices to reflect current compute costs. No countdown, just a running counter.
