Blazil — Open-Core Financial and AI Infrastructure
0TPS
10× Visa peak · fault-tolerant · failover tested live. AI-ready transport and verified distributed inference.
Blazil v0.3 was benchmarked on April 19, 2026 on an AWS i4i.4xlarge with 4 shards committing into a single 3-replica TigerBeetle VSR cluster. One replica was killed at t=80s — the cluster recovered in 37s with zero rejected events across 12,421,068 events. The same codebase now powers five production-grade dataset types, Tract ONNX inference, and a verified Qwen2.5-7B distributed pipeline. Real disk I/O, real network, real failover, and now a real AI infrastructure path on top of the same transport discipline.
Searching for Blazil? This is the official Blazil software website from Kolerr Lab Inc. Read the brand overview.
Production Benchmarks · v0.3 · April 19 2026
233K TPS. Zero errors. Live failover tested.
Blazil v0.3 was benchmarked on AWS i4i.4xlarge with 4 shards committing to a 3-replica TigerBeetle VSR cluster. One replica was killed at t=80s — the cluster recovered in 37 seconds with zero rejected events. This is our primary production-grade claim: fault-tolerant by design, failover live-tested.
Avg 180,500 TPS · 12,421,068 events processed · 0.000% error rate · VSR fault-tolerant consensus
$1.496/hr on i4i.4xlarge vs $5.29/hr TigerBeetle reference hardware — same VSR class, lower cost envelope
Killed replica at t=80s during live benchmark · cluster recovered with zero rejected events · real production test
How Blazil compares to industry leaders
vs Industry
Published peaks for competitors · Blazil v0.3 (233K TPS) is our primary production claim
Why this matters: Blazil achieves 9.7× Visa's published peak (24K TPS) and 46× Stripe's estimated throughput (~5K TPS) on cost-effective hardware. Our v0.3 benchmark ran for 120 seconds with live VSR failover testing — significantly more rigorous than typical industry reports.
Cost profile against TigerBeetle reference hardware
TigerBeetle (Published Benchmark)
$5.29/hr
- • Published TigerBeetle benchmark reference
- • i8g.16xlarge class hardware
- • ~38 second benchmark duration
- • No live failover test published
Blazil v0.3 (AWS i4i.4xlarge)
$1.496/hr
- • 1× AWS i4i.4xlarge instance
- • 16 vCPU (Intel Xeon Platinum 8375C)
- • $1,077/month at 24/7 utilization
- • 120 second sustained benchmark
- • Live VSR failover at t=80s (37s recovery)
3.5× lower hourly cost while running 3× longer with live fault tolerance testing. On-demand pricing for ap-southeast-1 region.
Architecture
From client to committed ledger without a single unnecessary copy.
gRPC streaming enters Go services, crosses into the Rust engine over TCP + MessagePack, lands on a lock-free LMAX Disruptor, and commits to TigerBeetle through VSR with O_DIRECT storage semantics. The same transport discipline now also underpins priority routing, cross-shard 2PC, and distributed AI inference.
01
gRPC Bidirectional Streaming
Persistent connections with 256 in-flight requests per stream eliminate TCP handshake overhead on every transaction. The channel stays open; only the data moves.
02
TCP + MessagePack Handoff
Go services forward validated traffic into the Rust engine over a compact framed transport so the control plane stays explicit and the hot path stays lean.
03
LMAX Disruptor Ring Buffer
A single-producer lock-free ring buffer processes events at 84ns P99 latency — faster than a CPU cache miss. No locks means no contention, and no contention means linear scaling.
04
Cross-Shard 2PC
TigerBeetle pending, post, and void flows enable atomic reserve-and-commit behavior across shards instead of leaving distributed transfer correctness to application code.
05
Priority Routing
Independent Aeron streams let critical events target <1ms latency while high and normal traffic keep their own backpressure and fairness guarantees.
06
TigerBeetle VSR + io_uring
Viewstamped Replication across three nodes provides fault tolerance while io_uring keeps the storage path tight enough for both fintech and AI workloads to share the same operating model.
Full Stack
AI Infrastructure
The same zero-copy discipline now powers datasets, inference, and distributed LLM serving.
Blazil is no longer just a high-throughput ledger engine. It now extends the same transport, observability, and operational model into AI workloads: Tract ONNX inference, io_uring data loading, five production-grade dataset types, and a verified Qwen2.5-7B distributed pipeline.
Pure Rust Inference
Tract ONNX inference and io_uring data loading remove Python and GPU runtime assumptions from the serving path.
Dataset-Agnostic Transport
All workloads flow through the same Sample<Vec<u8>> pipeline, so text, time-series, audio, and vision reuse the proven transport layer.
Distributed LLM Pipeline
Qwen2.5-7B is already running through a 3-stage Aeron IPC pipeline with token feedback, KV cache continuity, and multi-token generation.
Operator-Ready Foundations
Security, observability, and supply-chain controls are shared across fintech and AI workloads instead of being bolted on later.
Dataset Coverage
AI Metrics
Inference RPS
1,500–2,000
AWS i4i.4xlarge
Tract ONNX, CPU-only estimate pending production benchmark
Cost per RPS
$0.54–0.72/RPS/mo
AWS i4i.4xlarge
Comparable small-scale efficiency without GPU fleet overhead
Verified LLM run
32 tokens in 19.7s
Apple M4 CPU
Qwen2.5-7B-Instruct on distributed 3-stage pipeline
Why this matters
The README now positions Blazil as financial and AI infrastructure. The site needs to show that the AI side is already concrete: datasets are implemented, inference plumbing exists, and the distributed LLM path is verified rather than hypothetical.
Licensing
Open-core with a clear conversion path.
Blazil is source-available under Business Source License 1.1. Non-commercial use and research are free without restriction. Every release converts automatically to Apache 2.0 after four years. Production deployments require a commercial license.
Open Source
forever
- Full source access under BSL 1.1
- Free for non-commercial use and research
- Self-hosted evaluation with no time limit
- Community support through GitHub Issues
- Converts to Apache 2.0 — no action required
Commercial
for pricing
- Production deployment rights across all markets
- Commercial use license with legal clarity
- Priority support with an SLA guarantee
- Architecture review with Kolerr Lab Inc engineers
- Enterprise compliance and audit package
BSL 1.1 converts to Apache 2.0
Automatically, four years after each release date
Roadmap
Fintech scale, AI inference, and operating maturity now move on one roadmap.
v0.1 delivered 62,770 TPS on a 3-node DO cluster. v0.2 achieved 436K TPS sharded and 131K TPS VSR. v0.3 shipped April 19 2026 on AWS i4i.4xlarge — 233,894 TPS with live VSR failover testing. v0.3.1 integrated AI inference with Tract ONNX. v0.3.2 added multi-stream priority routing with <1ms critical latency. v0.3.3 hardened the platform with 2PC, mTLS, signed artifacts, and operator runbooks. v0.5 AI LLM verified distributed Qwen2.5-7B generation, and v0.5 AI Production now points toward large-model serving.
- —Core Rust engine with LMAX Disruptor ring buffer
- —TigerBeetle VSR consensus across three nodes
- —gRPC + Tokio UDP transport · 256 in-flight window
- —3-node DigitalOcean cluster · $252/month · 0% error rate
- —Aeron IPC transport — 1,203,108 TPS peak on MacBook Air M4
- —io_uring disk writes · sharded TigerBeetle E2E integration
- —2 shards per node · 436,351 TPS aggregate (3-node sharded)
- —130,998 TPS with full VSR consensus · 2/3 quorum · 0% error
- —DO cluster benchmark completed April 13 2026 · SGP1
- —Production-grade single-shard with live VSR failover test
- —AWS i4i.4xlarge · Intel Xeon Platinum 8375C · 16 vCPU · 128 GiB · 1.9TB NVMe
- —4 shards × dedicated TigerBeetle client · 3-node VSR cluster (loopback)
- —233,894 TPS peak · 180,500 TPS avg · 12,421,068 events · 0% error
- —VSR failover tested: 1 replica killed at t=80s → recovered in 37s · bench continued
- —AWS Singapore · April 19 2026
- —5 production-grade datasets implemented
- —Tract ONNX runtime integration for AI inference
- —io_uring dataloader for high-throughput model serving
- —2,291 LOC · 57 tests passing · CI 100% green
- —AI inference pipeline implemented · production benchmark pending
- —Multi-stream priority routing (Critical/High/Normal)
- —Critical requests bypass queue with <1ms latency guarantee
- —429 tests passing · 0 Clippy warnings
- —Production-ready priority scheduler deployed
- —Cross-shard 2PC via TigerBeetle pending/post/void flow
- —Kubernetes ingress + cert-manager mTLS automation
- —Prometheus PVC persistence and operational hardening
- —Syft SBOM + Cosign keyless signing in CI
- —3 ADRs and 8 runbooks published for operators
- —4× AWS i8g.16xlarge instances (Graviton 4, 64 vCPU each)
- —Sharded VSR cluster with horizontal scaling
- —Production benchmark pending · target 1M+ sustained TPS
- —Multi-region replication architecture
- —Qwen2.5-7B-Instruct distributed 3-stage inference pipeline
- —Stage orchestration across Aeron IPC streams 1001/2001/2002/1002/1003
- —Language Drift issue fixed and English generation verified
- —KV cache preserved across decode steps with correct position propagation
- —Production-ready multi-token generation on CPU-only Apple M4
- —Cortex v1 on ClarkenAI 70B plus Blazil Super Engine
- —Cloud cost validation against commercial inference APIs
- —Ankatos runtime integration and shared transport layer
- —Production benchmark pending for large-model serving
The source is open and the benchmarks are reproducible.
Clone the repository, run the benchmark suite, and verify every number yourself.
Enterprise
Talk to the Architect.
If you are evaluating payment ledger infrastructure, planning a core banking modernization, or building a low-latency AI serving stack on the same operational substrate, the right conversation is with the engineer who designed and built the system.
- 233,894 TPS · 1 VSR cluster · 4 shards · 12,421,068 events · 0% error
- 5 production-grade dataset types · Tract ONNX · distributed Qwen2.5-7B verified
- Cross-shard 2PC · mTLS hardening · SBOM + Cosign supply-chain controls
Direct contact
Ricky Anh Nguyen · Founder & Principal Architect
General · licensing · partnerships
No sales funnel. No SDR. A direct line to the engineer who built the system.