Enterprise AI often stalls at the chatbot stage: useful pilots, but fragmented tools, duplicated integrations, manual testing and limited access to trusted data. Azercell chose a different path, making AI a reusable, governed execution layer for the company. AI Battleground is the in-house platform built for this. Employees can use approved AI, build agents in Agent Builder, connect them to 40+ approved MCP services and share successful agents through an internal Marketplace. The MCP Gateway controls tool access and credentials, the LLM Gateway separates applications from model providers and supports routing and anonymization. Prompt and model changes can be evaluated against versioned scenarios before promotion. The same architecture can run cloud-native, hybrid or fully air-gapped with local models. Since launching in January 2026, Battleground has reached 337 users, 243 active users and 125 owned agents. It has recorded 28,553 agent calls, 17,142 MCP calls and 6,176 code executions, an internal proxy indicates that about 70% of active conversations go beyond plain chat. The platform already supports operational workflows: Meeting Agent turns Teams recordings into structured AZ,EN,RU summaries, AI Board Navigator supports CxO Business Reviews with historical context and meeting intelligence, Data Intelligence Agent lets authorized employees work with governed enterprise data in natural language. Battleground also connects enterprise adoption with local AI capability. With AWS, Azercell established an Azerbaijani model-training framework using custom tokenization, continued pre-training and LoRA fine-tuning, AWS reports about 2x better tokenizer efficiency. The result is not another AI pilot, but a foundation for scaling governed, reusable and locally relevant AI.

Nominasiya: İlin Özəl AI Həlli
Şirkət: Azercell Telecom LLC
Layihə: AI Battleground
Sahə: Creative AI / дизайн / видео / музыка
Mərhələ: Масштабирование
AI komponenti: AI is Battleground's execution layer. Agents combine LLM and SLM reasoning with enterprise knowledge, structured data, code execution and approved tools to complete multi-step tasks, not simply generate text. More than 40 MCP servers expose capabilities such as SAP, Jira, Starburst and Tableau through a governed MCP Gateway with RBAC, fine-grained permissions and credential isolation. The LLM Gateway decouples agents from model providers, enabling controlled routing, anonymization and tokenization and use of cloud or local models. Execution telemetry demonstrates this agentic use: 28,553 agent calls, 17,142 MCP calls and 6,176 code executions. An internal proxy estimates that about 70% of active conversations invoke an agent, enterprise tool or code capability. AI is also embedded in production workflows. Meeting Agent processes Teams recordings into structured AZ, EN and RU summaries, AI Board Navigator combines historical reviews, meeting intelligence and memory for CxO decision support, Data Intelligence Agent answers natural-language questions over governed enterprise data. Quality is part of the AI system itself. Prompt and model changes can trigger a version-controlled evaluation suite combining deterministic checks, such as guardrail adherence, groundedness and citation presence and latency, with rubric-based grading. Candidates are compared with production baselines, regressions can be blocked or routed to SME review, and production failures can become new test cases. Human accountability remains in place for consequential claims, recommendations and actions.
Təsir / nəticələr: Since its January 2026 launch, AI Battleground has moved from platform build to measurable production adoption. Current telemetry shows 337 users, 243 active users (73%), 125 owned agents, 6,564 conversations and 44,582 requests. The strongest signal is how the platform is used: 28,553 agent calls, 17,142 MCP calls and 6,176 code executions. An internal proxy indicates that about 70% of active conversations go beyond plain LLM chat and invoke an agent, enterprise tool or code capability. This activity supports reusable workflows. Meeting Agent automates the path from Teams recording or transcript to a structured AZ, EN and RU summary and participant delivery. AI Board Navigator supports CxO Business Reviews with pre-reads, multi-period context and meeting intelligence. Data Intelligence Agent lets authorized users ask business questions over governed enterprise data in natural language. Scale also comes from reuse: teams build on one platform, gateways, controls and evaluation framework instead of recreating separate AI stacks, while successful agents can be shared through the Marketplace. External evidence strengthens the case: AWS reports about 2x tokenizer efficiency, 23% higher training throughput and 58% lower peak GPU memory in the documented Azerbaijani setup, Starburst has published Azercell's production agentic use cases and unified-data work.
İnnovasiya: AI Battleground is differentiated by what surrounds the model. Instead of a licensed chatbot or isolated pilots, Azercell built the complete enterprise agent lifecycle as one governed platform. Employees create agents through Agent Builder, connect approved capabilities through 40+ MCP servers, validate changes automatically and publish successful agents to an internal Marketplace. This moves AI creation from a central specialist team toward controlled self-service. Governance is architectural. The MCP Gateway separates agents from backend credentials and enforces access controls. The LLM Gateway separates applications from model providers, enabling routing, anonymization and controlled model substitution. Quality is managed like software release engineering: prompt or model changes can trigger re-evaluation against a versioned corpus, regressions can block promotion, production failures can become new test cases. The platform can run cloud-native, hybrid or fully air-gapped with local models. For Azerbaijani, it goes beyond translation: custom tokenization, local SLM development and emerging ASR and TTS can be reused inside enterprise agents.
Azərbaycanla əlaqə: AI Battleground was designed and built in Azerbaijan by Azercell's in-house team, developing capability in agentic AI, MLOps and LLM operations. Adoption builds workforce capability: 337 users are on the platform and 125 agents have been created, moving employees from consuming AI toward building reusable solutions. The platform also addresses a local technology gap. Azerbaijani is a comparatively low-resource language, so Azercell worked with AWS on a repeatable training framework using a custom tokenizer, continued pre-training and LoRA fine-tuning. AWS reports about 2x better tokenizer efficiency. Azercell is extending this with Azerbaijani SLM, ASR and TTS. Battleground aligns with Azerbaijan's AI Strategy for 2025 to 2028 in infrastructure, Azerbaijani NLP, workforce skills, data governance, security and responsible AI, translating national priorities into private-sector capability.
Komanda: 6–15 человек
Biznes modeli: Внутреннее корпоративное внедрение
İşə başlama tarixi: 2026-01

Sayta keç

← Bütün nominantlar