§ 01Agentic Systems Engineering·Claude
We put engineers inside your team to get Claude into production — then leave behind an evaluation harness that shows exactly what each new model release changes for your workflows, so you can adopt it sooner.
Nineteen years of production infrastructure. Hadoop, then Kubernetes, now agentic systems. Different substrate, same discipline.
An evaluation suite of 7 cases run against 3 model versions. All cases passed on the earlier versions. On the newest version, 1 case fails — refund_authority — caught before merge.
1 regression · caught before merge
Every engagement ships one of these. It keeps running after we leave.
§ 02The practice
Agentic systems engineering on Claude — from whether the problem is tractable, through architecture and the build, to production. Evaluation is the last stage, not the offer.
Whether agents can do the job at all — including when the answer is no.
Orchestration, tool boundaries, state, and what happens when a step fails.
The build, as production code in your repo, through your review process.
Deployment, observability, security review, and a handover your team can hold.
Eval suites in CI, drift detection, and re-baselining on every model release.
Every decade, a new substrate becomes the next decade of enterprise infrastructure. VISystems has shipped at every inflection since 2007. The technology changes. The engineering discipline — eval-first, telemetered, audit-trailed, open-source — does not.
2007–2015
Built production Apache Big Data platforms when the stack was still in incubation. Contributed upstream, ran the infrastructure at F1000 scale, and learned a discipline: if you can’t evaluate it, you can’t ship it.
Built with: Apache Hadoop, Spark, Cassandra, HBase, ZooKeeper
2016–2023
Designed multi-cluster Kubernetes platforms, service-mesh topologies, and GitOps CI/CD for regulated financial services and large ISVs. Shipped internal tooling that became the engineering playbook for the next substrate.
Built with: Kubernetes, Istio, Argo, Flux, Prometheus, OPA
2024 → present
Current era. Designing and building agent systems that survive contact with production — orchestration, tool and MCP interfaces, context and cost engineering, then the eval harness that proves it still works on each new model release. 7 OSS repos published. The harness runs against Anthropic, OpenAI, Google, AWS Bedrock, and Azure OpenAI.
Shipping: Eval Cloud (SaaS), eval-runner, eval-as-ci-action, model-migration-kit + 4 more OSS repos
We work with enterprises whose compliance and procurement constraints mean we can’t name them here. The list below is anonymized to the level their legal teams are comfortable with.
§ 05Platform capabilities
Eval Cloud covers the full AI model lifecycle — from wiring up your first eval suite to producing the audit evidence that satisfies a model risk officer. Five core capabilities, four operational dimensions, multi-provider from the start.
§04.1
Eval suites, context architecture, tools, MCP servers, agents. Built test-first against a published spec, version-controlled in your repos, instrumented from day one.
§04.2
Production traces, observability, drift detection, capacity planning, eval-as-CI gates. The day-two work that determines whether agents stay in production a year from now.
§04.3
Automated migration safety on every model release. Audit ledger by default. SR 11-7 alignment documentation. The work that survives a model risk officer's review.
§04.4
Automated reporting. Cross-provider benchmarking. Your reliability posture improves every quarter — the platform surfaces what to optimize next.
§ 06Tiers — platform plans
Every tier ships every product feature — SSO, SCIM, the eval harness, MCP registry, audit ledger, all CI integrations, full data export. Tiers differ on eval volume, retention, and isolation — never on whether you can do something. No SSO tax. No SCIM tax. No audit-ledger tax.
| Feature | §05.1 Free Get started | §05.2 Team Production scale | §05.3 Business Enterprise-ready · SR 11-7-aligned | §05.4 Enterprise Custom contracts · dedicated tenancy |
|---|---|---|---|---|
| Eval runs / month | 100 | 5,000 | Unlimited | Unlimited |
| Models monitored | 1 | 5 | Unlimited | Unlimited |
| Providers | 1 | 3 | All | All |
| Migration lead time | Standard | Priority | Next business day | Same day SLA |
| CS cadence | Quarterly | Quarterly + monthly | + named CSM | Dedicated CSM |
| Trace retention * | 12 mo | 24 mo | 84 months · SR 11-7 | 84 months · SR 11-7 |
| DB tenancy * | Shared | Shared | Dedicated Neon project | Dedicated · custom region |
| Specific capacity for the Enterprise tier is scoped on request. Early-access subscribers who join before public beta lock preferential pricing and free historical-telemetry migration at GA. | ||||
* Eval Cloud platform features — currently in private beta. Available on every tier at GA.
→ Full pricing detail at /pricing
§ 07Platform capabilities · five core modules
Eval Cloud ships every capability to every subscriber. Eval execution, drift detection, migration safety, cost intelligence, and compliance evidence are platform features — not add-ons, not tier-gated modules. You connect your providers, configure your gates, and the platform runs continuously.
BUILD · DEFINE SPECS, CONNECT PROVIDERS, GATE CI
RUN · REAL-TIME MODEL BEHAVIOR MONITORING
RELY · AUTO-REBASELINE ON EVERY MODEL RELEASE
RUN · CROSS-PROVIDER COST TRACKING AND OPTIMIZATION
RELY · AUDIT TRAILS, SR 11-7 DOCUMENTATION
→ Full platform documentation with coverage details and integration guides at /offerings
§ 08 Eval Cloud · the operational layer
▌ Built because we needed it · available on every tier
Every AI deployment needs evals, model-upgrade review, and evidence that survives audit. Eval Cloud turns eval and migration history into marketplace-procured model-risk evidence: run suites, detect provider releases, re-baseline behavior, flag regressions, and produce the audit trail regulated teams need.
Currently in PRIVATE BETA. Public beta Q3 2026. v1 GA target Q1 2027.
▌ Private beta — closed
We are working with a first cohort and are not taking new sign-ups. General availability is targeted for Q1 2027.
If you carry a regulated workload and want to talk about it early, dev@visystems.com.
§ 09 Three differentiators
We bet our brand on three load-bearing differentiators.
▌ Differentiator 01
Eval Cloud instruments token spend, latency cost-curves, and model substitution opportunities across Anthropic, OpenAI, Google, AWS Bedrock, and Azure OpenAI from the moment you connect. We publish the metrics — anonymized and aggregated — in the VISystems Impact Report so subscribers can benchmark their own cost trajectory against the cohort. If your cost-per-eval is drifting, the platform flags it before you notice it in billing.
→ VISystems Impact Report · published quarterly to subscribers
▌ Differentiator 02
7 Apache 2.0 repos, all built by the same engineers who build the platform. There is no separate OSS team — the engineers who build the platform write the OSS. The repos are the proof that the discipline we ship is the discipline we practice.
→ Q1 · eval-runner · live
▌ Differentiator 03
Published platform with documented capabilities. Every feature has documented coverage and integration guides. No ambiguity about what you get. Start free, upgrade when your workload grows, export your data anytime.
→ Self-serve · 5 CI integrations · data portability