Security and compliance

Security, compliance and sovereignty

Deployment options, the compliance matrix, model governance, data residency and what is shared in a discovery session.

Updated: September 2026

Deployment options

AI must meet the same standards as every other critical system. Aginies is built to the requirements of regulated sectors such as finance, insurance and healthcare; data stays where it must. Four deployment models deliver the same platform.

OptionWhere data livesWhere models runWhen it fits
On-premise / air-gapThe enterprise’s own data centre; no external connectionoss and speech models served insideThe most sensitive data; regulatory obligation
HybridSensitive workloads insideoss inside; outside only with masked data through a secure gatewayElastic capacity needs; a mixed scenario portfolio
Private cloud / dedicated regionA single-tenant region reserved for the enterpriseInside the region; data residency guaranteedEnterprises that want residency guarantees without running a data centre
GPU-as-a-ServiceInside national bordersElastic GPU capacity that scales with usageOn-prem model serving without capital outlay

Compliance matrix

Aginies is a SOC 2 Type II and ISO 27001 compliant platform; the control set is documented for internal audit. The matrix below summarises the compliance areas and how the platform addresses each.

AreaStatusWhat it covers
SOC 2 Type IICompliant platformThe control set is documented for internal audit
ISO 27001Compliant platformInformation security management
KVKK & GDPRCompliancePersonal data map, data residency guarantee, deletion and access requests
BDDK IT regulationCompliant deploymentFor banks, primary and secondary systems in-country; audit-ready deployment
SSO / SAML & RBACBuilt inFine-grained roles, workspace isolation, per-agent access policies
PII masking & anonymisationBuilt inMasking at ingestion, configurable redaction, anonymisation of all requests
Audit trailBuilt inInput, output, model version, rule set and user for every agent action and model call
Kill-switchBuilt inEmergency shutdown and automation limits on critical decisions; mandatory human approval
“Compliant platform” means the control set is designed and documented against the relevant standards; independent certification is a separate process and is documented separately.

Model governance

Model-agnostic, fine-tuned. Hub Aginies holds enterprise fine-tuned versions of proprietary models, open-source-based on-prem models and Turkish speech models in one catalogue. There is no vendor lock-in; switching models does not mean rewriting the agent.

  • Model proxy: cloud models reached with the enterprise’s own keys through one point
  • Keys stored encrypted; usage reported per agent, unit and channel
  • Every request to an external model is masked and anonymised
  • Fallback model strategy: operations continue through provider outages
  • Prompt versioning: which agent ran with which prompt and model version is recorded
  • On-prem oss models for sensitive scenarios; data never leaves the enterprise

Data residency and sovereignty

Sovereign design means control of data and models stays with the enterprise. In on-premise and air-gapped deployments no data leaves; in hybrid deployments only masked data goes outside, through a secure gateway. GPU-as-a-Service provides elastic capacity inside national borders.

  • Personal data map: which data, in which system, for what purpose, for how long
  • Data residency guarantee: contractual in dedicated-region and on-prem options
  • For banks, primary and secondary systems in-country
  • Deletion and access requests: recording and response aligned with KVKK and GDPR
  • Voice and code data: Voice and Code Aginies run on-prem

The code-assistant scenario: from IDE to model

For the information-security teams of organisations that want to work in a closed circuit. The organisation’s developers reach Hub Aginies through the Code Aginies extension in their IDEs, over the organisation’s own VPN, on our infrastructure in data centres located in Türkiye and aligned with banking IT regulation; the coding models run there, on in-country GPUs. Code context never leaves Türkiye, no external model is called, no model is trained on the organisation’s data, and every request leaves a per-user trace.

ORGANISATION NETWORK · ON-SITE Developer IDEsVS Code · JetBrains · organisation device and user Code Aginies extension signed · IDE Registry organisation licence key + SSO user identity Organisation identity providerSSO · SAML / OIDC · MFA · group-based authorisation Organisation SIEM / log collectionoptional · trace forwarding via syslog / API VPN TUNNEL org. gateway ↔ Aginies gateway IPsec · IP allowlist in-country link AGINIES SOVEREIGN PERIMETER · TÜRKİYE · REGULATED DATA CENTRE Hub Aginies · API gatewaysingle entry point · Platform Aginies module authentication licence · quota rate limit PII / secret masking routing · fallback tenant isolation Licence serviceorg. key · seats · term Platform Aginiesobservability and governance audit trail (per request) metrics · latency · errors usage · cost kill-switch · RBAC Model servers · GPU · in-countrycoding models selected from the Hub catalogue aginies-tulpar-4 aginies-sunkar-1 fine-tuned families on Qwen · GLM · Kimi · versioned PUBLIC INTERNET / CLOUD MODEL CALLS: NONE TRAINING ON ORGANISATION DATA: NONE BACKUPS OR LOGS OUTSIDE TÜRKİYE: NONE data centres: Türkiye · regulation-aligned © Aginies · conceptual architecture 1 SSO session · MFA 2 code context + prompt · TLS 1.3 suggestion returns the same way 3 4 masked prompt 5 suggestion · response 6 7 trace to the organisation’s SIEM · optional · one-way
  1. Identity: The developer signs in from the IDE with the organisation’s identity provider (SAML / OIDC, MFA). The extension carries a short-lived per-user token; the licence key is held at organisation level and stays outside the extension.
  2. Request: The extension sends only the relevant code context (open file, cursor surroundings, selected symbol definitions) and the prompt; the repository as a whole is never sent. The connection is encrypted with TLS 1.3 inside the organisation’s VPN tunnel; the Hub accepts traffic only from the IP addresses allocated to the organisation.
  3. Licence and quota: Hub Aginies validates the token, binds it to the organisation’s workspace and checks seats and quota. When a user is disabled at the organisation, access drops at the same moment.
  4. Masking and routing: The prompt is scanned with configurable rules for secrets (keys, passwords, connection strings) and personal data; matches are masked. The prompt is routed to the model selected in the contract.
  5. Model: The model runs on GPUs in Aginies’ data centre in Türkiye. There is no external API call; prompt and response are not stored on the model side and are not used for training.
  6. Response and trace: The suggestion returns to the IDE through the same tunnel. Platform Aginies records the request with user, time, model version, duration, tokens and cost; whether prompt and response content is retained, and for how long, is decided together with the organisation.
  7. The organisation’s visibility: The organisation’s administrators see usage and cost dashboards; traces can be forwarded one-way to the organisation’s SIEM via syslog or API. The kill-switch cuts access instantly per user, team or organisation.
LayerControlEvidence / audit
IdentitySSO with the organisation’s IdP (SAML / OIDC), MFA at the organisation; user- and group-level authorisation; instant revocationSession and authorisation logs; seat list with the organisation’s administrator
NetworkTwo access models: in the VPN model a site-to-site tunnel and a Hub unreachable from outside; in the internet model TLS 1.3 and the organisation’s egress IPs on an allowlist (rate-limit and auto-block exemption)Tunnel or allowlist configuration; connectivity tests; network diagram with IP addresses
ApplicationSigned extension, approved-version distribution from the IDE Registry; version pinningSignature verification; release notes; SBOM on request
DataSecret and PII masking; context limiting; encryption (TLS in transit, AES-256 at rest); retention period by contractMasking rule set; retention policy; deletion evidence
ModelServed on in-country GPUs; no external calls; no training; versioned models, validated with a test corpus, fallback modelModel catalogue and version history; test reports
TenantThe organisation’s workspace is logically isolated; dedicated tenant or dedicated hardware on requestIsolation design; access matrix
ObservabilityPer-request audit trail; usage and cost dashboards; SIEM forwarding; kill-switchDashboards; log samples; forwarding integration
OperationsIncident notification, change management, backup and disaster recovery inside TürkiyeIncident response plan; RTO/RPO; drill records
ComplianceSOC 2 Type II and ISO 27001 compliant control set; KVKK; deployment aligned with banking IT regulationControl set document; penetration-test summary; audit right in the contract

Two access models: VPN tunnel or allowlist over the internet. The same request path opens in two ways. A VPN tunnel when the organisation wants a fully closed circuit; TLS over the internet with the organisation’s egress IPs on an allowlist when developers should start without installing anything extra. The second model was set up together with a bank’s information-security team; the diagram and flow table below are the anonymised version of that deployment.

  • VPN tunnel (fully closed circuit): Site-to-site IPsec between the organisation’s gateway and Aginies; the Hub is reached only through the tunnel from the organisation’s allocated IPs and is invisible from the internet.
  • Over the internet, with an allowlist (quick start): The endpoints are reachable from the public internet; the extension connects over TLS 1.3 from the organisation’s existing internet egress. The organisation’s egress IPs go on an allowlist so that rate-limit and auto-block rules do not cut its traffic; access itself is granted by identity.
ORGANISATION NETWORK · ON-SITE Developer workstationsIDE + Code Aginies extension · SSO, MFA code context + prompt only · repo stays local no VPN client or server to install Organisation internet egressproxy / NAT / firewall · existing egress egress IPs → Aginies allowlist single outbound rule: Aginies public IP · TCP 443 © Aginies · anonymised deployment INTERNET 1 HTTPS · 443 · TLS 1.3 no VPN · allowlist same-session reply AGINIES SOVEREIGN PERIMETER · TÜRKİYE · REGULATED DATA CENTRE Aginies application serversingle public IP · every service on TCP 443 code-api.aginies.comextension API hub-api.aginies.comdirect model access hub.aginies.comHub interface app · dialog.aginies.comPlatform and Dialog rate limit / auto-block: organisation IPs exempt SSO token · seats · quota · masking · per-request trace SSH 22 · Aginies operations VPN only · closed to the organisation External model providers0 calls · no training ✕ Aginies operationsVPN → SSH · change management 4 INTERNAL NETWORK · /24 · NO INTERNET EGRESS 2 model request · internal network 3 response the same way GPU model serverno public IP · reached only from the application server coding models on in-country GPUs · versioned · fallback model prompt and response not stored · not used for training
FlowSourceTargetPort / protocolAccess and control
1Developer IDE · Code Aginies extension · organisation egress IPscode-api.aginies.com → Aginies public IPTCP 443 · HTTPS, TLS 1.3Internet; organisation IPs exempt from rate limit and auto-block. SSO token, seats and quota, masking, per-request trace.
2Aginies application server (internal interface)GPU model server (internal network)Internal network · model APIInternal network only; accepted only from the application server, never leaves the data centre.
3GPU model serverInternet—None: the server has no public IP and no internet egress; prompt and response are not stored or used for training.
4Aginies operations teamAginies application server · administrationTCP 22 · SSHAginies operations VPN only; closed to the internet and to the organisation. No SSH or VPN needed on the organisation side.
5Aginies application serverExternal model providers—None: coding models run only on in-country GPUs; code context never leaves Türkiye.
6Organisation applications (optional, direct model access)hub-api.aginies.com → Aginies public IPTCP 443 · HTTPS, TLS 1.3Same allowlist, identity and quota; API access to the models in the Hub catalogue.

Addresses in the diagram are anonymised; the real IPs, hostnames and firewall rules are given to the organisation’s information-security team in the technical annex.

Location and infrastructure

Where do the systems run? Does data leave Türkiye?

The whole platform runs in our data centres in Türkiye; primary and secondary systems are in-country. Code context, logs and backups never leave Türkiye. This is aligned with the in-country requirement of banking IT regulation.

Which data centre, and to which standards?

The systems run in data centres located in Türkiye that hold international Tier and national TS EN 50600 certifications and are aligned with financial-sector regulation (banking and payments regulators). The data centre’s name, documents and certificates are shared in the discovery session and in writing in the contract annex.

A dedicated environment or a shared one?

The default is a logically isolated workspace for the organisation on sovereign shared infrastructure. At the organisation’s choice a dedicated tenant, a dedicated GPU pool or an on-prem deployment in the organisation’s own data centre is possible; the architecture is the same in all three.

Who are the sub-processors? Do you use a cloud provider?

No cloud provider is used for model serving. Data-centre services come from providers located in Türkiye and approved by the regulators; the provider is named in the contract and the organisation is informed in advance of any change.

Network and connectivity

How is the connection set up? Which ports and addresses?

There are two options. In the VPN model a site-to-site IPsec tunnel is established between the organisation’s gateway and Aginies, only HTTPS (443) is used inside the tunnel and the Hub accepts traffic only from the IP addresses allocated to the organisation. In the internet model the endpoints are reachable over TLS 1.3 from the public internet; the organisation opens a single outbound rule (Aginies public IP, TCP 443) and its egress IPs are added to an allowlist. The allowlist keeps rate-limit and auto-block rules from cutting the organisation’s traffic; access itself is granted by identity. In both models the IP addresses and parameters are given in the technical annex.

Is Hub Aginies exposed to the internet?

No. The Hub is reachable only through the tunnel interface; it has no public endpoint, so there is no DDoS or scanning surface.

How is encryption handled?

TLS 1.3 in transit inside the VPN tunnel (end-to-end, independent of the tunnel); AES-256 at rest. Key management is on the Aginies side; bring-your-own-key is available in the dedicated deployment.

Can the organisation’s proxy or SSL inspection sit in between?

The extension honours corporate proxy settings. TLS inspection with the organisation’s own certificate is technically possible; whether the inspection device sees code context is then the organisation’s own policy. Our recommendation is to keep end-to-end TLS inside the tunnel.

Identity and access

How are users authenticated? What if the licence key is shared?

Identity lives in the organisation’s identity provider: SSO via SAML or OIDC, MFA by the organisation’s policy. The licence key is organisation-level and grants no access on its own; every request carries a short-lived per-user token. When a user is disabled at the organisation, Aginies access drops at the same moment.

Who can use which model, and how much?

Authorisation is group-based: which team reaches which model, with which quota and which features (completion, chat, agents); the organisation’s administrator manages it from the Platform Aginies dashboard. Seats and quotas are set by contract.

Can Aginies staff access the organisation’s data?

Production access is role-based, named and logged; access to prompt and response content is opened only on the organisation’s written request for support purposes, and the record is shared with the organisation. Under the default retention policy, content is not kept at all.

Do administrative actions leave a trace?

Yes. Administrative actions are written to the same audit trail: who changed what, and when. Policy changes are versioned.

Data flow and retention

What exactly does the extension send?

Only the context the request needs: the relevant part of the open file, cursor surroundings, selected symbol definitions and the developer’s prompt. The repository as a whole, build outputs and environment files are never sent. The context limit and the exclusion list (.env, certificates, key files) are configured with the organisation and visible in the extension.

Do secrets (passwords, API keys) reach the model?

The masking layer in the Hub scans the prompt for keys, passwords, connection strings and personal-data patterns; matches are masked before they reach the model. The rule set can be extended with the organisation’s own patterns; masking events appear in the trace.

Are prompts and responses retained? For how long?

Metadata (user, time, model version, duration, tokens, cost) is always recorded. Retaining prompt and response content is the organisation’s decision: off (default), or encrypted retention for a period the organisation sets. Deletion at expiry is automatic and evidenced.

Is our data used to train models?

No. The organisation’s prompts and responses never enter any model’s training or fine-tuning data; this is an explicit clause in the contract. If organisation-specific fine-tuning is requested, a separate model version owned by the organisation is produced only from data the organisation approves.

Where are backups, and who can access them?

Backups are kept inside Türkiye, encrypted and under a separate access policy; if content retention is off, backups contain no content either. Backup retention and disaster-recovery targets are written in the contract annex.

What happens when the contract ends?

The organisation’s workspace, configuration and traces are exported in the requested format; all data is then deleted within the agreed period and deletion evidence is provided. An organisation-specific model version is handed over on request, otherwise destroyed.

Model and AI

Which models are used? Does anything go to an outside provider?

Coding models from the Hub Aginies catalogue served on in-country GPUs: open-source-based on-prem models (aginies-talay-2, aginies-sunkar-1) and fine-tuned enterprise versions. No prompt goes to an outside provider; the cloud model proxy is off in the organisation’s workspace and can be locked at policy level.

Who selects the models, and how are they updated?

Models enter the catalogue after Aginies reviews them and validates them with a test corpus; every model is versioned. A model change in the organisation’s workspace happens with the organisation’s approval and prior notice; the previous version is kept for rollback for an agreed period.

Is model output safe? Can it produce harmful code?

Model output is a suggestion; the developer reviews the code and it passes the organisation’s existing CI, code review and static analysis. On the Aginies side, check families for prompt injection and off-policy output run in the test corpus; organisation-specific forbidden patterns can be added as policy.

Licensing and IP: who owns the generated code?

The generated code belongs to the organisation. The licences and commercial-use terms of the open-source models in use are listed in the contract annex.

Extension and client side

Where is the extension installed from, and is it signed?

The extension is signed and distributed as an approved version through the organisation’s enterprise IDE Registry; no public marketplace is needed. Versions can be pinned; updates are released with the organisation’s approval.

Does the extension connect anywhere else? Is there telemetry?

The extension connects only to the Hub address the organisation configures; there is no other endpoint. Error-report telemetry is minimal, carries no content and can be switched off.

What if the connection drops? Does it work offline?

Without a connection the extension goes quiet; it keeps no local queue or cache, so no code context accumulates on the device. It resumes when the connection returns.

Can we review the source code or an SBOM?

The extension’s software bill of materials (SBOM) and dependency scan are shared on request; a source-code review can be arranged under an NDA.

Compliance, audit and operations

Which certifications do you hold?

The platform is compliant with the SOC 2 Type II and ISO 27001 control sets; the control-set document is shared for internal audit. The KVKK data-processor agreement and privacy notices are ready.

For banks: is this an outsourced support service under banking regulation? Is there an audit right?

If the organisation is a bank, the service may fall under its outsourced-support classification; the required contract clauses (audit right, sub-processor notification, exit plan, information-security obligations, audit access for the regulator and the organisation) are then added. The in-country requirement is already met by the architecture.

Are penetration tests performed? Can we see the results?

Independent penetration tests are performed regularly; the executive summary is shared and finding-closure evidence is provided. The organisation may run its own test; scope and timing are agreed in advance.

What is the notification time for a security incident?

For an incident affecting the organisation, first notification is made within 24 hours, followed by a root-cause and corrective-action report. KVKK notification duties are coordinated with the organisation as data controller.

Service level and capacity?

Availability and latency targets are written in the contract; capacity is reserved by the organisation’s seat count and the GPU pool scales in-country. Maintenance windows are announced in advance.

Can we cut the service instantly in an emergency?

Yes. The kill-switch is in the organisation administrator’s hands: it cuts access instantly per user, team or organisation; the organisation can also close the VPN tunnel on its own side to end all access.

How does change management work?

Every change to the Hub, the extension or the models that affects the organisation is announced in advance with release notes; critical changes can be tied to the organisation’s approval. Policy and configuration changes are versioned in the audit trail.

The same architecture applies to three deployments: sovereign shared infrastructure, dedicated tenant / GPU, on-prem in the organisation’s data centre.

What is shared in a discovery session

A discovery session is a working meeting to understand the enterprise’s scenarios and constraints. The following materials are shared so that security and compliance teams get their answers early.

  • Reference architecture and a detailed schema of the deployment options
  • Documentation of the SOC 2 Type II and ISO 27001 compliant control set for internal audit
  • Data flow and masking design: which data reaches which model, in what form
  • Model catalogue and proxy setup; key management and fallback strategy
  • A sample audit report and run trace
  • The five-layer assessment scorecard for the selected scenarios
If the enterprise’s sector-specific requirements are shared before the session, the materials are prepared for those constraints.

Sector regulatory frames

The regulatory frame and recommended deployment model from each industry page. The first-90-days plan and measured targets are on the industry page.

SectorFrameRecommended deploymentFirst measured target
BankingBanking IT and electronic banking regulation: primary and secondary systems in Türkiye, audit right in the outsourcing contract, cloud model subject to bank-specific permission · AML and personal-data law: suspicious-transaction reports decided by analysts, masking and purpose limitation in every scenario · Per-request audit trail for internal audit and regulator review: user, model version, data source, reasoningStart with the sovereign shared deployment in our data centres in Türkiye (isolated bank workspace, VPN or allowlist access); on-prem in the bank’s data centre for scenarios classed as primary systems.Simple transactions reaching the call centre: 40% fewer
Fintech & PaymentsCentral-bank payment and e-money IT communiqué: community cloud only with an approved provider; primary systems in-country · AML obligations: transaction-monitoring alerts to the analyst with evidence, the reporting decision with people · Card-scheme rules and personal-data law: chargeback deadlines tracked, card data masked and never used beyond purposeIn the community cloud of a provider on the central bank’s approved list, or in the institution’s own data centre; GPU capacity in-country, models from the Hub catalogue.Merchant request first-contact resolution: 70% and above
InsuranceInsurance regulator rules and insurance IT requirements: claim and policy decisions reasoned, time-stamped, open to audit · Special-category personal data: health and damage photos processed inside the perimeter, masking and access logging mandatory · Payment and rejection decisions always with the authorised adjuster or underwriter; the agent prepares the fileDedicated region in Türkiye or the sovereign shared deployment; vision models for damage photos and health data on in-country GPUs.Simple claim closure time: From weeks to days
Retail & E-commerceConsumer law and distance-selling rules: withdrawal, return and delivery deadlines coded into agent rules · Personal-data law and marketing consent: contact only on consented channels, recommendations on purpose-limited data · Campaign and price rules: discount authority bounded by the matrix, out-of-band offers to customer serviceSovereign shared deployment; GPU capacity scales on usage at seasonal peaks, customer data stays in Türkiye.First-contact resolution: +15–25 points
TelecomTelecom regulator rules and the electronic communications law: subscription, contract and number-porting operations recorded and traceable · Personal-data law and traffic data: subscriber data in Türkiye, recordings and transcripts masked · SIM swap and identity operations: extra verification and human approval; suspicious requests to the fraud teamSovereign shared deployment or a dedicated tenant in the operator’s own data centre; speech models on in-country GPUs, connected to the existing switch over SIP.Requests resolved without a representative: 60–70%
HealthcarePersonal health-data regulation and special-category data: health data inside the hospital perimeter, access by role and purpose, every access logged · Ministry hospital-information and national health-record integration rules: the agent acts through the HIS and keeps no separate data copy · Clinical decisions with the physician: the agent prepares, queries and organises; physician approval card when in doubtOn-prem in the hospital data centre; models inside, no external connection. For hospital groups, a dedicated tenant in the group data centre.No-show rate: 30–40% lower
Manufacturing & AutomotiveOT security (IEC 62443 approach): agents connect to the production network read-only and issue no commands; OT/IT separation kept · Product safety and warranty law: warranty decisions rule-based, out-of-tolerance files to people · Personal-data law and supplier contracts: supplier and customer data inside the organisation, sharing limited to the contract scopeHybrid: read-only agents and historian access near the production network on site, decision and document agents in our data centre in Türkiye or the organisation’s own.Unplanned downtime: 20–30% lower
LogisticsCustoms law: declaration and document preparation by the agent, declaration and signature with the authorised broker · GDP and cold-chain rules: temperature excursions documented, the decision with the quality officer · Personal-data law and customer contracts: shipment and consignee data inside the organisation, every customer message recordedSovereign shared deployment; TMS, telematics and carrier APIs connected through an allowlist, shipment data stays in Türkiye.Delay detection lead: Hours ahead
Energy & UtilitiesEnergy regulator consumer-service and connection regulations: dispute, correction and application deadlines coded into agent rules · Critical infrastructure and OT separation: read-only access to SCADA and OMS; grid decisions in the control room · Personal-data law and meter data: consumption data only for the customer’s own account, retention by policySovereign shared deployment for customer operations; grid-adjacent read-only agents on-prem in the company data centre with critical-infrastructure separation kept.Calls resolved without a representative: 55–65%
Government & Public SectorPresidential information and communication security guide: deployment in the institution’s data centre, air-gap included; open-weight models inside · Administrative procedure and public procurement law: every output with the regulation article and source document; the decision with the authorised officer or commission · Personal-data law and e-Government integration rules: citizen data purpose-limited, access by authorisation, full audit trailOn-prem in the institution’s own data centre; air-gapped where required. For municipalities and affiliated bodies, a dedicated tenant in the public cloud.Documents asked of the citizen again: 70% fewer
Defence & AerospaceDefence-industry agency and export-control law: control-list mapping by the agent, licence and sharing decisions with the authorised compliance officer · Facility security clearance and classification levels: access separation, marking checks and an audit log of every access · ITAR/EAR-like restrictions in allied programmes: controlled technical data only to cleared users, no egress to external networksAir-gapped deployment: models, platform and data inside the closed network; updates as approved packages, remote maintenance only in windows the organisation opens.Party screening time: From days to minutes
IoT & Connected DevicesPersonal-data law and device data: telemetry inside the organisation, user data purpose-limited · Product safety and cyber-resilience requirements (incl. EU CRA for exporters): OTA packages signed, rollout recorded, rollback ready · Safety-critical devices: read-only; commands and bulk operations under product-owner approvalHybrid: device cloud and time-series data on the existing infrastructure, agents and small models in our data centre in Türkiye or the manufacturer’s own.Faults needing a field visit: 35% fewer
Software DeliverySource code and intellectual property: code never leaves the organisation; Code Aginies and completion models on-prem · Secure development requirements (incl. regulator expectations in regulated sectors): SAST, dependency and secret scanning on every PR, evidence stored per release · Production deployment and architecture decisions with people; the agent writes only in assigned repositories and branchesCode Aginies and coding models in the organisation’s data centre or in our data centre in Türkiye through the closed-circuit deployment; SDLC agents connected to Jira, Git and CI/CD with service accounts.PRs accepted on first review: 60–75%