Platform documentation

A · Orchestration & Execution Intelligence

The engine layer. Runs complex work end-to-end; routes requests intelligently and keeps multi-step processes stable even when dependencies fail. Every run is stateful: steps, intermediate results and decisions are recorded.

• Supervisor → sub-agent → supervisor architecture
• Dynamic agent routing
• Stateful multi-step execution; parallel branches and loops
• Agent-to-agent communication (A2A)
• Scheduled and webhook triggers
• Agent templates and the visual agent builder
• Autonomous Agent: build agents by dialogue with test corpora

B · AI / Model Orchestration

Brain selection and resilience. Uses the right model for each task; reduces vendor lock-in and keeps operations running when a provider is unavailable. Switching models does not require redesigning the agent.

• Multi-model gateway: proprietary and open-source models behind one interface
• Model proxy: observable, masked access to cloud models with your own keys
• Model switching without redesign
• Fallback model strategy
• On-prem serving for open-source models
• Prompt testing environment and versioning
• Hub Aginies: fine-tuned model families on Qwen, GLM and Kimi bases, current open releases and speech models

C · Knowledge & Retrieval

Enterprise memory. Turns fragmented enterprise information into governed, searchable, trustworthy knowledge; answers link back to sources. Chunk-level control makes RAG quality directly manageable.

• Document embedding pipeline
• Versioned knowledge base management
• Tag-based document segmentation
• Hybrid search and multi-source RAG
• Chunk-level inspection and editing
• Hallucination-reduction architecture
• Source, tag and access-policy management

D · Governance, Risk & Compliance

The control layer. Makes AI auditable, safe and fit for regulated environments; every decision is traceable. Governance is not bolted on; it is part of the architecture.

• Guardrails and policy engine
• Human-in-the-loop checkpoints: threshold and approver based
• Agent decision traceability
• Sensitive data masking and PII redaction
• Masking and anonymisation of all requests
• Full input/output logging and audit trail
• Kill-switch and automation limits

E · Observability & Optimisation

Performance management. Provides visibility into accuracy, speed, cost and health; produces the data for continuous improvement. Every run is recorded step by step.

• Per-run traces and timelines
• Latency and token tracking
• Cost attribution by scenario, unit and channel
• Usage and cost per model; platform health
• Routing accuracy measurement
• Dashboard and analytics layer
• Live logs, conversation records, per-agent success heatmap

F · Integration & Connectivity

Enterprise reach. Connects to enterprise systems and customer channels without rebuilding the existing technology landscape. Each tool carries its own invocation policy.

• REST API and database integrations
• Model Context Protocol (MCP) servers
• Custom tool definitions and invocation policies
• Native connectors: email, calendar, files, collaboration tools
• Web widget and messaging integrations
• MCP Hub: catalogue of approved tool servers

G · Multichannel Experience & Voice

The customer interaction layer. Delivers AI across digital, messaging and voice channels on one consistent orchestration underneath. The same agent runs in text and voice.

• Real-time speech-to-text and streaming voice processing
• Call-centre SIP integration
• WhatsApp, SMS and email services
• Intent detection and entity extraction blocks
• Session memory
• Dialog Aginies: dialogue UI and human takeover
• Voice Aginies: Turkish STT, TTS and ASR models

H · Deployment & Enterprise Architecture

Fit to enterprise constraints. Supports the on-premise, hybrid and sovereignty requirements of large enterprises. Data stays where it must.

• On-premise and air-gapped deployment
• Hybrid architecture: sensitive workloads inside, elastic capacity outside
• Private cloud / dedicated region
• SSO / SAML, fine-grained RBAC, workspace isolation
• Per-tenant agent sharing and collaborative development
• GPU-as-a-Service for in-country elastic capacity

Autonomous Agent

The Autonomous Agent builds agents by dialogue. You describe what you want; it sets up the supervisor and specialist agents, wires tools and the knowledge base, chooses models, defines approval points, writes and runs a test corpus, and deploys after your approval. Every agent in the library can be produced this way.

1. Describe: objective, channels and approval rule are given in natural language.
2. Build: supervisor and specialist agents are created; a version number is assigned.
3. Wire: tools (REST/MCP) and the knowledge base are attached; chunk counts are reported.
4. Model selection: separate models for intent, decisions and speech; on-prem preferences respected.
5. Approval points: threshold and approver defined (e.g. amount > 100,000 → branch manager).
6. A test corpus is written and run; check families are scored.
7. Human approval: results are presented; the deployment decision always stays with a person.
8. Deploy: the new version goes live; logs and dashboard are active immediately.

A test corpus has three levels: suites (golden path, branches, robustness injection), scenarios within each suite, and multi-turn conversations within each scenario. Every turn is scored against check families.

• customer_resolved: was the customer’s request actually resolved
• pii_before_verification: was personal data disclosed before identity verification
• State machine: did the conversation stay within defined states and transitions
• Memory round-trip: was information from earlier turns recalled correctly
• Judge dimensions: qualitative criteria such as accuracy, tone and policy adherence
• Tool and approval checks: right tool, right parameters, approval triggered above threshold

Coverage shows what share of defined intents, branches and tools the corpus exercises. Low coverage is surfaced as a warning before deployment; results are kept per version and compared across versions.

Knowledge Base

The knowledge base lets agents answer with sources. Documents pass through the embedding pipeline, are split into chunks, tagged and versioned. Each agent knows which sources it may reach, with which tags and under which access policy.

• Sources: files, folders, web, databases and connectors
• Chunk editor: inspect, edit or disable any chunk
• Versioning: which agent ran with which knowledge base version is recorded
• Hybrid search: semantic and keyword retrieval combined
• Access policy: visibility by tag and role
• Citation: every claim in an answer links to its chunk

Hub Aginies: Model Hub, MCP Hub and proxy

The model catalogue holds three groups in one place: Aginies families fine-tuned on enterprise data from Qwen, GLM and Kimi bases (tulpar, talay, sunkar, konrul); the current open releases of the same bases; Turkish speech and vision models. Family names are stable, version numbers move: when a new base release ships, the fine-tune is repeated and the catalogue version increments; the previous version stays available for the contract term. Each scenario gets the model that balances accuracy, latency and cost. The versions below are the October 2026 catalogue.

Model proxy: you enter your own cloud keys and the platform reaches cloud models on your behalf as a proxy. Every call passes through one point, is observed and masked. Keys are stored encrypted; usage and cost are reported per agent, unit and channel.

• Switch models with one selection; the fallback strategy takes over during provider outages
• Prompt testing environment and versioning; comparison across versions
• MCP Hub: catalogue of approved tool servers; access and invocation policy per server

Voice Aginies

Speech models trained for Turkish: streaming speech-to-text (STT), natural text-to-speech (TTS) and real-time ASR. Voice agents connect through call-centre SIP integration and run on the same orchestration as text-channel agents.

• Models: aginies-voice-tr-v3 (STT), aginies-tts-tr-v2, aginies-asr-stream-v1
• On-prem serving; audio never leaves the enterprise
• Voice analysis: quality scoring, sentiment and intent
• 100% of calls transcribed and scored against the quality form
• Personal data is masked during transcription

Vision Aginies

Image and document understanding: invoice, contract and form extraction, damage-photo assessment, visual quality control. Personal-data detection and document masking run on the same models, inside the enterprise.

• aginies-konrul-1 · on-prem
• Field extraction with source document and confidence score per field
• Document classification, independent of layout
• Personal-data detection and document masking (KVKK/GDPR)
• Low-confidence fields route to human verification

Code Aginies

For engineering teams: IDE extensions and in-house code completion, an enterprise IDE registry that distributes approved extensions and models, and the SDLC agent team working in the CI/CD pipeline. Code never leaves the enterprise.

• In-house code completion on on-prem models
• IDE Registry: approved extension and model distribution, version control
• SDLC agent team: PM, analyst, architect, developer, tester
• Works with Jira, Git and CI; code review and the production decision stay with people

Dialog Aginies

The dialogue surface for customers and employees: web widget, mobile, WhatsApp and messaging. Session memory, human takeover and an audit trail for every message. The same agent runs in text and voice channels.

• Human takeover and hand-back; context handed to the representative with a summary
• Conversation dashboard and quality scoring
• Intent detection and entity extraction blocks
• Integration with the identity flow; unauthorised requests stop at the guardrail

Observability and reporting

What happens per agent, how the platform is doing and what each model costs: on one dashboard. Every run is recorded step by step; logs and conversation records are reviewed under role-based access.

• Live runs: step, duration, tokens, target system and outcome
• Per-agent success heatmap and error timeline
• Quality and error timelines; routing accuracy
• Cost attribution by scenario, unit, channel and branch
• Usage and cost per model; platform health
• Capacity projection: cost and resource forecast with usage growth
• Audit reporting: input, output, model version, rule set, user

Reports can be exported and retained in a form suitable for internal audit and regulatory requests. Conversation records are stored masked; unmasked access requires separate authorisation.

Workspaces, tenants and collaboration

The platform is separated by workspace and tenant. Each team manages its own agents, knowledge bases and keys in isolation; approved agents are shared inside the organisation and developed together.

• Identity through SSO / SAML; permissions through fine-grained RBAC
• Workspace isolation: data, logs and cost reported separately
• Per-tenant agent sharing and collaborative development
• Templates, users and licences under central management
• Agent versions and release history recorded per workspace

Platform facts derived from the library

The figures and distributions on the platform page are derived from the agent library at build time. The same data is given here as tables.

Integration catalogue

The 200 systems the scenarios touch, in 12 kinds. Connecting a new system means one tool definition and one invocation policy; write tools require approval by default.

Operating model

The platform’s standard operating model after go-live. Targets are settled with the organisation in the contract.

Product screens

Platform Aginies · Dashboard · last 30 daysall workspaces · all tenants

executions13,120,000

success97.4%

p50 latency1,4 s

platform healthhealthy

busiest agents · each cell ≈ 1 day

dialog_banking_assistant%97.4

card_debt_payment%99.1

kvkk_data_detection%96.2

kvkk_document_masking%98.8

motor_claim_end_to_end%98.1

merchant_operations%100

model usage

aginies-tulpar-438%

aginies-talay-229%

aginies-sunkar-119%

aginies-voice-tr-v39%

cloud · proxy$2.960

cost · by scenario

Dialogue banking26%

Credit & onboarding24%

Claims operations18%

KVKK masking14%

Merchant ops10%

Other8%

PII masking: onanonymisation: onaudit trail: full1,920+ active agents

Platform Aginies · Reportsall workspaces · October 2026

monthly executions · last 12 monthslast 30 days: 13,120,000+31% YoY

report	breakdown	schedule	format
Monthly cost report	scenario · unit · channel · branch	1st of month · 07:00	PDF · XLSX
Agent performance	success, latency, error classes	weekly · Monday	PDF
Audit export	run traces, approvals, model versions	on demand	CSV · JSON
Model usage	tokens, cost and proxy share per model	daily · 06:00	XLSX
SLA and capacity	p95 latency, queues, GPU occupancy, projection	weekly	PDF

delivered by email and shared folderinternal audit templatepersonal data masked

Dialog Aginies · web widgetcard_debt_payment v2

I want to pay my full card balance from my current account.

Your statement balance is 24,860, due 14 October. Current account ****7741 has enough balance. If you confirm I will send a one-time code.

✓ get_card_debt · get_accounts · policy: no approval needed

Confirm.

Code sent. Once entered, the payment executes instantly and you get a confirmation.

session memoryhuman takeover: readyaudit trail

Type a message…↑

Voice Aginies · SIP · call 0212-…-4471aginies-voice-tr-v3 · on-prem

customerSomeone hit my car last night, plate 34 ABC 123, I want to report a claim.

agentSorry to hear that. I found your policy by plate; can I take the location and time?

STT180 ms

intentclaim intake · 0.97

sentimentcalm

TTS · aginies-tts-tr-v2ASR · stream-v1PII masking: on

Vision Aginies · invoice_2026_1012.pdfaginies-konrul-1 · on-prem

extracted fields

Invoice noFTR-2026-1012

Date12.10.2026

KDV%20 · 3.125,00

Total18.750,00 TL

TCKNmasked (KVKK)

source: page 1confidence ≥ 0.95

Code Aginies · billing-service/invoices.tscompletion · on-prem

14export async function exportInvoicePdf(id: string) {
15  const invoice = await repo.findById(id);
16  if (!invoice) throw new NotFound('invoice');
17  const pdf = await renderer.render('invoice', maskPii(invoice));
18  await audit.log('invoice.export', { id, user: ctx.user });
19  return storage.put(`invoices/${id}.pdf`, pdf);
20}

Tab: acceptPROJ-4821coverage 87%SAST ✓

IDE RegistryVS Code · JetBrains extensionDeveloper Agent · PR #2291

Hub Aginies · Model Huball workspaces · October 2026

model	access	usage · last 30 days	cost
aginies-tulpar-4	on-prem · GPU 4×	58%	—
aginies-talay-2	on-prem · GPU 2×	84%	—
aginies-sunkar-1	on-prem · GPU 1×	91%	—
aginies-konrul-1	on-prem · vision	22%	—
qwen · current release	on-prem · open model	31%	—
aginies-voice-tr-v3	on-prem · STT	47%	—
cloud · proxy	your key	12%	$288

keys encrypted · never shown to usersall requests maskedMCP Hub · 14 servers

Workforce Aginies · Roster and work logworkspace: fintech-ops · this week · tenants separate

agent	role · team	hour-eq.	output	quality	mode
ios-dev-01	iOS Developer · mobile team	14 h	9 PRs · 21 points	4.7	autonomous · 10% review
qa-01	QA Engineer · mobile team	31 h	412 tests · 6 bugs	4.8	autonomous · 10% review
ba-01	Business Analyst · payments product	9 h	11 stories · 2 processes	4.5	shadow · 100% review
acct-01	Accountant · finance	27 h	238 entries · 3 reconciliations	4.9	autonomous · threshold-approved

work log · today

09:12ios-dev-01PR #4821 opened: payment screen accessibility fixes · 2 tests addedreview pending

09:40acct-0138 supplier invoices processed, 2 out of tolerance → to m.ozkan for approvalapproval

10:05qa-01Regression round 3: 412/412 · 1 new bug (BUG-2291) filed with evidencedone

10:30ba-01Story PAY-118 with 6 acceptance criteria written → to the product ownerreview

weekly timesheet exportedbilling: agent-monthteam lead: e.demir (Aginies)