pilot · June 2026logistics AI assistant in operation — architecture, tool calling, agentic RAG, MCP server, evals
9 → 65agent-driven analytics metrics in one month; verified against the reference report, ≤3% tolerance
1 sprinttrading module for logistics marketplaces — idea to production (Python, backend + frontend)
−80%release defects after changes to planning and code review
week → 1 daytime-to-market after moving to a release-free delivery model
3 days → 30 minanalytics turnaround on ClickHouse / DataLens
Logistics AI assistant at Kamaz Digital
Users are internal employees; the key working scenarios serve freight-forwarding managers and logisticians. The agent operates on top of corporate systems (ERP, TMS, internal APIs). Pilot in operation since June 2026.
- Problem
- Answering routine employee questions required manual lookups across several corporate systems.
- What I did
- Designed the agent architecture and led early development: scenarios, skills and prompts, tool calling, agentic RAG (query expansion, agentic retrieval); designed the knowledge layer (entity and procedural knowledge) and the analytics layer on top of the corporate DWH.
- Result
- The assistant has been pilot in operation since June 2026 (internal users); the effort grew into a dedicated track that keeps adding new scenarios on top of this architecture.
employee request
│
▼
scenarios (intents) ──▶ tool calling ──────▶ agentic RAG ──────▶ answer
ERP · TMS query expansion │
internal APIs agentic retrieval │
▲ │
└────── evals gate: scenario dataset · quality · run cost ──────┘Assistant DWH analytics: 9 → 65 indicators in a month
A semantic layer on top of the corporate DWH (ClickHouse). A business-driven initiative; iterative changes were made by a coding agent.
- Problem
- Business questions ran into manual SQL and number reconciliation: the assistant could compute 9 canonical indicators, and scaling was blocked by trust in the numbers.
- What I did
- Built a closed loop around a reference — the recurring management Excel report with key business indicators. Correctness criterion: the layer reproduces the report's numbers on historical data within a ±3% tolerance. A coding agent edited the semantic layer and replayed all references against the live DWH until convergence.
- Result
- In one month the layer grew from 9 to 65 verified indicators (×7). Of 27 layer versions only 5 were bug fixes: the loop mostly produced new indicators rather than repairing old ones.
management Excel report (key business indicators)
│
▼
references: metric · period · value · tolerance ±3%
│
▼
coding agent edits the semantic layer ──▶ replay on historical DWH data
▲ │ mismatch with a reference
└──── all references converge → layer release ────┘
Globaltruck IT: −80% release defects, delivery week → day
A carrier's logistics platform. Product team of 12, backend team of 3 developers.
- Problem
- Releases shipped once a week as heavy batches and regularly carried defects.
- What I did
- Rebuilt planning, code review and retrospectives; moved the team to a release-free model — features deploy as they are ready.
- Result
- Release defects dropped by 80%; changes reach production in a day instead of a week.
Kamaz Digital data platform: analytics in 30 minutes instead of 3 days
10+ data sources, reporting for the group's operations management.
- Problem
- Every business analytics request was assembled by hand over 3 days.
- What I did
- Designed a ClickHouse platform, integrated 10+ sources, built data marts and DataLens reporting.
- Result
- A typical report is ready in 30 minutes; the AI assistant later grew on top of this platform.
