Building AI
systems that
actually run
in production.

Engineering teams waste hours searching scattered Confluence spaces for answers. KnowledgeBot brings AI-powered search directly into Slack — scoped to what each user is permitted to see, with citations back to source pages.

• Multi-tenant vector store (per-space ChromaDB collections) with Slack group-based access control
• HyDE search — generates hypothetical document fragments for dramatically better embedding matches
• Zero-downtime atomic reindexing — builds replacement collection in background, swaps in one operation
• Jira task creation with duplicate detection via vector similarity + LLM assessment
• Documentation quality audit: 6 parallel checkers, LLM-powered auto-fix suggestions
• Per-feature API cost attribution with Slack overspend alerts

Fully stateless design — all persistent state lives in Jira and GitLab. This enables safe restarts and horizontal scaling without a database. Claude Code runs as a subprocess (not SDK) to enable tool restrictions, credential isolation, and MCP server injection per invocation.

• End-to-end autonomous dev: Open task → plan → code → CI → review → merge
• Multi-layer resilience: global + per-group circuit breakers, exponential cooldown, pipeline retry with CI log analysis
• 21-state configurable workflow with human approval gates via stop-labels
• Comment-driven directives — users inject mid-workflow instructions via @mention
• Release window enforcement: day/hour/timezone config + holiday API integration

Custom metaclass-based ORM with thread-safe in-memory caching and temporal query support. ML-driven inventory forecasting using 168-bin hourly profiles (hour × weekday), monthly seasonality, and IQR outlier filtering — predicts depletion up to 30 days ahead.

• 7 commission types across partner hierarchies with retroactive recalculation
• CAS-based distributed task chain tracking via Redis — beyond Celery's native primitives
• Bidirectional 1C ERP sync: payments, agreements, invoices, tax agent data
• 16+ report types (PDF/Excel/CSV) with auto-scheduled monthly partner distribution
• Multi-channel alert escalation: Slack → Telegram → SMS → Email with Redis cooldowns

Clean Architecture with composable 17-state FSM — regional subclasses define base states, product flows inject their states at runtime. LLM Provider abstraction supports Claude, GPT-4o, and Gemini with automatic fallback chaining and per-call cost metering.

• AI-driven sales conversation — Claude collects, negotiates, and extracts structured contract data from natural language
• Automated KYC with risk scoring (0–1) via government business registries — auto-approve or route to moderation
• Multi-region contract generation via Jinja2 templates with EDO electronic signing
• Automated widget deployment verification: HTTP checks + HTML parsing + JS loading validation
• Redis distributed locking prevents race conditions from rapid multi-message input