AI Employee Reporting & Executive Intelligence System

The Project Story

The Challenge

Organizations faced a dual bottleneck in operational human resources:

• Bottom-Up Bottleneck: Employees submitted vague or incomplete daily activity reports, requiring managers to manually review, follow up, and verify details before authorizing daily clock-outs.
• Top-Down Blindspots: Directors and C-level executives lacked centralized real-time visibility into weekly KPI progress, team obstacles, and individual employee performance without convening repetitive daily briefing meetings.

The Solution: Multi-Platform Hybrid AI Architecture

I engineered an end-to-end Conversational HR & Intelligence Ecosystem deployed across dedicated communication channels on self-hosted infrastructure. Rather than relying on a single monolithic prompt, the system decouples operational employee intake from executive decision intelligence while sharing a unified Supabase PostgreSQL memory layer.

The Business Impact

Daily reporting compliance and detail quality reached 99% across operational teams. Shift clock-outs are approved automatically without managerial micro-management, while executive leadership receives actionable daily intelligence and instant analytical Q&A access — cutting daily status meetings and completely eliminating third-party HR SaaS subscription overhead.

System & Interface Showcase

Interactive screenshot gallery showing the conversational multi-platform AI agents, internal reporting database, and integration portals.

Architecture Overview

A hybrid microservices architecture synchronizing event-driven workflow automation with a custom Python AI analytical engine across multi-channel interfaces.

*Note: Database schema names, table entities, and internal identifiers in this case study have been generalized and sanitized for confidentiality.

• Multi-Platform Messaging Gateways: Self-hosted WhatsApp gateways (WAHA) for operational employee intake, with support for Telegram Bot APIs for executive analytical interaction.
• Operational Ingestion Layer: n8n handles event webhooks, dynamic multi-agent prompt routing, and attendance API triggers.
• Analytical Intelligence Service: Custom FastAPI backend processing scheduled morning executive briefings, Pydantic structured output validation, multi-table queries, and intent classification.
• Centralized Memory Layer: Supabase PostgreSQL consolidating operational schemas (employee_master, activity_reports, management_directives, chat_history, executive_summaries).
• Enterprise Portal: Web dashboard for HR administrators to search historical logs, inspect evaluations, and manage staff records.

Unified System Flow

[ Employee Intake Layer ]
Employees ──(WhatsApp)──► [ Gateway / Webhook ] ──► [ n8n Multi-Agent Workflow ]
                                                          │
                                                          ▼ (Pass / Reject & Clock-out Unlock)
                                              ┌──────────────────────────────────────────────┐
                                              │  Supabase PostgreSQL Data Layer              │
                                              │  (employee_master, activity_reports,         │
                                              │   management_directives, executive_insights) │
                                              └──────────────────────┬───────────────────────┘
                                                                     │
[ Executive Intelligence Layer ]                                     ▼
Directors ◄──(Telegram / WA)── [ Channel Gateway ] ◄── [ FastAPI Executive Intelligence Engine ]
          ──(Q&A & Directives)──►                  ◄── [ n8n Scheduled Cron (06:00 AM) ]
                               (Strict Auth &
                                Memory Engine)
                    

Agentic System & Memory Design

• Event-Sourced Management Memory: Preserves managerial context across sessions. Director directives are persisted, soft-deleted upon obsolescence, and factored into subsequent daily evaluations.
• Structured Pydantic Extraction: The intelligence engine outputs strictly validated JSON schemas, ensuring executive summaries and risk metrics remain consistent and audit-ready.
• Multi-Turn Conversational Window: The executive assistant maintains a rolling 10-message conversational context, enabling natural pronoun resolution ("what were his main blockers today?") during deep-dive sessions.
• Automated Attendance Gating: The operational interviewer validates reports against quality criteria before dispatching cryptographically signed authorization payloads to the attendance portal.

LLM Evaluation & Production Observability

To guarantee high reasoning accuracy, eliminate hallucinations in executive briefings, and maintain strict consistency, the system implements a dual evaluation strategy:

• Offline Evaluation (G-Eval & Golden Dataset): Before deploying prompt iterations or model updates, pipeline outputs are benchmarked against curated golden datasets using G-Eval (LLM-as-a-Judge). This systematically scores criteria such as factual alignment, report completeness, actionable summary synthesis, and adherence to Pydantic schemas.
• Online Observability & Tracing (Langfuse): Production conversations and analytical runs are continuously traced via Langfuse. This enables real-time monitoring of latency, token expenditures, LLM call traces, intent classification accuracy, and user feedback in live production environments.

AI Infrastructure & LLM Routing

Powered by 9router to ensure high availability, model resilience, and cost efficiency:

• Multi-Model Fallback: Automatically shifts payload traffic to secondary LLM endpoints if primary provider latency spikes or encounters rate limits.
• Tiered Inference Routing: Employs compact models for fast message classification and conversational acknowledgments, while routing multi-report synthesis and strategic evaluation to high-reasoning models.
• Containerized Deployment: Production instances packaged via Docker Compose with dedicated health checks and isolated networking.