Skip to content
ForgeStack AI

Additional Engineering Services

From evidence to engineering action

ForgeStack's primary services establish what is happening technically: Release Evidence shows what happened during production delivery, while an AI Provider Exit Drill tests what happens when a production AI workload must operate through a different provider or model arrangement.

Where those engagements expose a material gap, ForgeStack can scope focused engineering work to fix, automate, optimise or control it.

These are separate engineering engagements. They are not automatically included in Release Evidence or an AI Provider Exit Drill.

Each engagement establishes or improves the capability; the monthly service keeps it operating, tested or evidenced.

Evidence Automation

Turn recurring evidence reconstruction into an engineered process.

Keeps connected evidence sources collecting and reports completeness as systems change.

From £600/month, based on three connected evidence sources.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, recurring requests for release evidence send engineers back to pipeline runs, pull requests, approvals, security checks, deployments and spreadsheets to rebuild the same record. ForgeStack captures release metadata, links records across tools and produces a consistent evidence trail. After, an illustrative automated evidence flow shows sources connected, evidence generated, exceptions flagged and the record ready.BeforeRecurring evidence work“Can you produce the same release evidenceagain, quickly and consistently?”Pipeline runsPull requestsApprovalsSecurity checksDeploymentsSpreadsheetsRepeated reconstruction. Inconsistent evidence. Lostengineering time.Evidence AutomationCaptures release metadataLinks records across toolsProduces a consistent evidence trailAfterAutomated evidence flowIllustrativeRelease·Date·StatusSources connectedEvidence generatedExceptions flaggedRecord readyRepeatable evidence generation. Less manual work.More consistency.

Where release evidence repeatedly has to be gathered from pipelines, pull requests, approvals, testing, security checks and deployment records, ForgeStack can automate the collection and structuring of that evidence.

Typical work can include:

  • capturing release and deployment metadata;
  • linking change, approval and testing records;
  • identifying missing evidence;
  • recording exceptions;
  • producing consistent technical evidence records;
  • preserving the evidence trail around each release.

The objective is not another dashboard. It is a more reliable evidence-generation process using the systems the customer already operates.

Release Control Engineering

Move from recording controls to enforcing them consistently.

Release Control Operations

Chases every bypass, override and emergency change to closure, with a monthly control-health report and a quarterly control pack.

From £1,500/month per production pipeline.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, approvals, security gates, change references, tests, emergency releases and exceptions exist on paper, but some paths bypass them. ForgeStack implements agreed technical controls that warn, require approval or block, and records exceptions and outcomes. After, an illustrative controlled release path shows the approval check, security gate, change reference, exception path and recorded decision.BeforeControl and enforcementWe require approvals and checks, but can weprove the path actually enforces them?ApprovalsSecuritygateChangereferenceTestsEmergencyreleaseExceptionsPolicy on paper. Inconsistent enforcement.Release-control risk.Release Control EngineeringImplements agreed technical controlsWarns, requires approval or blocksRecords exceptions and outcomesAfterControlled release pathIllustrativePath·Environment·StatusApproval checkRequired and enforcedSecurity gateAutomated checks enforcedChange referenceLinked and validatedException pathRequires approval andjustificationDecision recordedAll outcomes capturedDefined controls. Visible exceptions. Consistentenforcement.

Evidence may show that required checks exist but can be bypassed, inconsistently applied or difficult to verify. ForgeStack can engineer customer-defined release controls around the existing delivery process.

Depending on the environment, this can include:

  • approval requirements;
  • security and test gates;
  • change-reference validation;
  • exception handling;
  • emergency-release paths;
  • evidence-completeness checks;
  • warning, approval or blocking behaviour determined by customer policy.

ForgeStack does not define the customer's governance policy. We implement the technical controls agreed with the customer.

AI Exit Remediation

Fix the portability failures exposed by an AI Provider Exit Drill.

From £5,000.

Continuity Plus

Ongoing continuity plus bounded remediation for provider and model change events found through the continuity cycle.

From £2,500/month per workload.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, an exit drill has found that on the fallback tool calls fail, schemas differ, retrieval drifts, permissions mismatch and latency is too high. ForgeStack fixes the drill findings, aligns prompts, tools and schemas, and re-tests after changes. After, an illustrative remediation status shows tool calls, retrieval, permissions and degraded mode working on re-test, with structured outputs still partial.BeforeDrill findingsThe fallback answered, but parts of theworkflow still broke. What has to be fixed?Your AI workloadTool calls failSchemas differRetrieval driftsPermissionsmismatchLatency too highKnown portability gaps. Unreliable fallback. Residualcontinuity risk.AI Exit RemediationFixes drill findingsAligns prompts, tools and schemasRe-tests after changesAfterRemediation statusIllustrativeDate·Re-tested·StatusAreaBefore (drill)After(remediation)StatusTool callsStructuredoutputsRetrievalPermissionsDegradedmodeRemediation notesTargeted fixes applied. Reduced risk. Re-test toconfirm completeness.

An exit drill can reveal that an alternative provider responds successfully while parts of the business process still fail. ForgeStack can remediate the technical issues identified during the drill, including:

  • provider-specific prompts;
  • incompatible tool calls;
  • structured-output differences;
  • retrieval behaviour;
  • context assumptions;
  • permission dependencies;
  • refusal and safety behaviour;
  • degraded-mode behaviour.

The repaired workload can then be re-tested against the agreed exit scenarios.

AI Model & API Cost Optimisation

Reduce AI operating cost without silently breaking the workload.

AI Cost Watch

Re-runs the cost-and-quality baseline every month, with one optimisation cycle each quarter.

From £1,000/month; price varies with monthly AI spend.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, an AI workload uses a premium model everywhere, with token waste, latency spikes, no routing logic and no optimisation baseline. ForgeStack measures cost, latency and behaviour, routes work to the right model and tests changes against the workload. After, an illustrative optimisation record compares the current path with the optimised path for model mix, cost, latency and acceptable quality, and lists the key findings.BeforeCost and performanceWhere is the spend, and how do we reduce itwithout breaking the workload?Your AI workloadPremium modeleverywhereToken wasteLatency spikesNo routing logicNo optimisationbaselineRising AI spend. Unmeasured trade-offs.Optimisation risk.AI Model & API Cost OptimisationMeasures cost, latency and behaviourRoutes work to the right modelTests changes against the workloadAfterOptimisation recordIllustrativeDate·Workload·StatusCurrent path(Before)Optimised path(After)Model mixCost(per 1M tokens)LatencyQuality acceptableKey findingsRoutes tasks to the mostcost-effective model per use caseReduces token waste with targetedprompt and output controlsMaintains quality within definedthresholdsLower spend. Controlled quality. Evidence-backedoptimisation.

The cheapest model or provider is not necessarily the cheapest way to complete a business process successfully. ForgeStack can use the workload's evaluation baseline to assess cost, latency and behaviour together.

Engineering work can include:

  • model selection;
  • workload routing;
  • prompt and context optimisation;
  • caching;
  • batching;
  • retrieval and embedding optimisation;
  • expensive-model escalation rules;
  • provider and model comparison;
  • inference right-sizing.

Changes are evaluated against the workload rather than judged on token price alone. The objective is lower operating cost while preserving the behaviour the customer actually needs.

Customer-Controlled AI Fallback

Build an AI fallback capability the customer controls.

Fallback Operations

Keeps the fallback patched, tests it every month and fails over to it each quarter, with a dated record.

From £2,500/month per environment. Business hours only.

Initial engineering scope quoted separately.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, the workload relies on one external provider, with no customer-hosted path, secrets outside the customer's control and no tested degraded mode. ForgeStack builds an alternative AI path on customer-controlled infrastructure and tests it against the workload. After, an illustrative fallback architecture shows the primary provider and a customer-controlled fallback joined by a runtime switch, with the customer's own retrieval, tools, secrets and models.BeforeDependency and controlWe rely on one external provider. What is ourcontrolled fallback?External providerSingle providerNo customer-hosted pathSecrets outside controlNo tested degraded modeSingle-provider dependence. Limited control.Continuity exposure.Customer-Controlled AI FallbackBuilds an alternative AI pathUses customer-controlled infrastructureTests it against the workloadAfterFallback architectureIllustrativePath·Environment·StatusPrimary providerExternal provider (for performance)Runtime switchCustomer-controlled fallbackYour infrastructure (under your control)RetrievalYour data andindexesToolsYour integrations andAPIsSecretsYour keymanagementModels/runtimeYour hosted runtimeand modelsA controlled fallback exists. Ready to test whenneeded.

Where reliance on a single external provider creates an unacceptable dependency, ForgeStack can engineer an alternative model path inside infrastructure controlled by the customer.

Depending on requirements, this can include:

  • open-weight or alternative models;
  • AWS- or Azure-hosted inference;
  • customer-controlled networking and secrets;
  • retrieval and knowledge sources;
  • application integrations;
  • tool and API connectivity;
  • fallback routing;
  • degraded operating modes;
  • workload evaluation.

Where the evidence supports it, model adaptation or fine-tuning can be considered as part of the implementation rather than assumed in advance. The result is tested against the customer's workload before being treated as a credible fallback.

AI Runtime Control Engineering

Make model choice, fallback and AI behaviour governable in production.

Runtime Control Operations

Keeps approved models, routing, budgets and fallback order current, with a monthly decision report.

From £1,500/month per gateway.

Onboarding may be required before the monthly service begins.

Starting price for the stated scope. Final scope and price are confirmed before work begins. Customer cloud, model and third-party licence costs are separate.

Before ForgeStack, user requests reach several models with ad hoc routing, an unclear fallback order, no schema validation, cost spikes and invisible exceptions. ForgeStack implements routing and guardrail logic, applies policy at runtime and records model and fallback decisions. After, an illustrative runtime control policy covers approved models, fallback order, cost threshold, schema check, a routing record and an exception log.BeforeRuntime governanceCan we see and control how models arerouted, validated and failed over?User requestModel AModel BModel CAd hocroutingUnclearfallback orderNo schemavalidationCost spikesExceptionsinvisibleUncontrolled runtime behaviour. Hidden risk. Weakgovernance.AI Runtime Control EngineeringImplements routing and guardrail logicApplies policy at runtimeRecords model and fallback decisionsAfterRuntime control policyIllustrativeVersion·Environment·StatusApproved modelsPrimary modelSecondary modelFallback modelFallback order1Primary model2Secondary model3Fallback modelCost thresholdMax cost per requestAction on breachSchema checkInput schemaOutput schemaValidation actionRouting recordTimeModelReasonStatusException logTimeTypeModelActionGoverned operation. Visible routing. Enforcedruntime rules.

Once a workload can operate across more than one model or provider, the next challenge is controlling how those options are used. ForgeStack can engineer runtime controls around areas such as:

  • approved providers and models;
  • model-routing rules;
  • schema and tool-call validation;
  • cost thresholds;
  • context limits;
  • fallback order;
  • provider-failure behaviour;
  • regional or hosting constraints;
  • degraded mode;
  • exception handling;
  • technical evidence of routing and fallback decisions.

Customer policy determines what should warn, require approval, fail over or stop. The objective is controlled operation rather than invisible model switching.

Commercial boundaries

  • business hours only;
  • no unlimited engineering or remediation;
  • you keep production approval and control;
  • significant new engineering is scoped separately;
  • cloud, model and licence costs are paid by you;
  • technical evidence and engineering, not legal advice or compliance certification.

Cost Watch, Fallback Operations and Runtime Control Operations can be bought without AI Continuity.

Where ForgeStack did not build or test the underlying system, onboarding or baseline work may be needed first.

Free AI Provider Readiness Check

Send one document, redacted export or short questionnaire about one production AI workload. We'll identify what is documented, what is exposed to change and what would need to be tested before it could be relied upon.

No system access. No production changes. No obligation.

The free check reads documents and exports only. It does not run an exit drill, test alternatives or change anything.

Request a Free AI Provider Readiness Check

One engineering chain

These services can connect naturally without requiring a customer to buy them all.

  1. Release Evidence
  2. Evidence Automation
  3. Release Control Engineering
  1. AI Provider Exit Drill
  2. AI Exit Remediation
  3. AI Model & API Cost Optimisation
  4. Customer-Controlled AI Fallback
  5. AI Runtime Control Engineering

The appropriate next step depends on what the evidence actually shows.

Discuss an engineering requirement

If an existing ForgeStack engagement has identified something that needs remediation, optimisation or automation, or your team already has a specific technical requirement in one of these areas, contact ForgeStack to discuss the scope.

Discuss an engineering requirement