Production AI System · Case Study

Executive Reporting AI Agent

An autonomous agent that ingests healthcare data, detects anomalies, and delivers tailored briefings to leadership before 7 AM. Every weekday. Unattended.

Built by
Alex Ovchinnikov
Roles LinkedIn →
Consulting solution4you.pro →
In Production Runs Every Weekday HIPAA Compliant · PHI Protected ~$1/day Operating Cost

The morning report writes itself before anyone logs in.

No dashboard to open. No analyst to prep it. The numbers that matter, the anomalies worth knowing, and the actions to take, delivered to the right inbox at the right altitude.

0%
Manual reporting overhead eliminated
~2 FTE
Analyst Time Recaptured
40→5%
AI Error Rate Reduced
~$0/day
Total Operating Cost
0
Power BI Seats Freed
7 AM
Insights Delivered Daily
0%
Autonomous · Zero Touch
See it work · live demo
Watch the agent produce a morning briefingEach run reflects a different day. Two streams: one for execs, one for ops.
01
Poll
02
Validate
03
Read 7d
04
Analyze
05
Ground
06
Deliver
Want a briefing engine like this for your team?
Built and deployed by Alex Ovchinnikov. Production agents, not prototypes.
Visit solution4you.pro →
Before
Manual Power BI dashboard. Analyst prep, technical navigation, daily refreshes. KPI breaches surfaced in meetings, not before them.
→
After
Fully autonomous push-model. The agent ingests, detects, and delivers to execs and ops leaders before 7 AM. Unattended, every weekday.
How it works · the engine
01🏥
Source
Client Systems
Healthcare transactional data
02🗄️
Infrastructure
AWS Databases
Production + staging pipeline
03⚙️
Data Engineering
Trim · Obfuscate · Aggregate
Python ETL with completeness checks
04📂
Central Hub
SharePoint
Daily files + KPI baseline targets
05🔗
Orchestration
AI Agent Platform
Polls, validates, reads 7-day window
06🧠
Intelligence
LLM Analysis
Won eval vs 8 models on quality + cost
07🛡️
Validation
Grounding Layer
Baseline anchoring + field-level checks
08📧
Output
2 Email Streams
Exec summary + ops action alerts
09👥
Delivery
Stakeholders
C-suite + ops leads, before 7 AM
Who it serves · business impact
👔
C-Suite
Strategic Awareness
  • ✓Daily summary in inbox by 7 AM — no dashboard login required
  • ✓Anomalies surfaced automatically — not manually hunted
  • ✓Decisions informed before the workday — trusted cadence
📋
Ops Leaders
Proactive Response
  • ✓Alerts only on threshold breaches — zero noise
  • ✓Prep before exec meetings — not scrambling during
  • ✓Clear accountability — each alert points to an area
🏢
Organization
Operational ROI
  • ✓70% manual overhead eliminated — analysts do higher-value work
  • ✓~$1/day fully loaded — replaces hours of analyst time
  • ✓Power BI dependency removed for daily exec decisions
🏥
Source Systems & AWS
Pre-existing Infrastructure
  • Sources Healthcare client transactional systems
  • Prod DB AWS-hosted production database
  • Staging AWS staging for analytical workloads
  • Ownership Maintained by existing teams
⚙️
Data Engineering Pipeline
ETL + File Generation
  • Tech Python + cron on AWS Lambda/EC2
  • Process Trim → Obfuscate → Aggregate
  • Validation Completeness checks before file gen
  • Output Dated daily Excel (.xlsx) to SharePoint
Staging → Trim → Obfuscate → Aggregate → SharePoint
📂
SharePoint — Central Hub
File Storage & Baselines
  • Daily Files Dated Excel files, accumulating
  • Baselines Persistent KPI targets, set with ops leaders
  • Review Quarterly cadence for baseline updates
  • History Agent reads current + 6 prior days
🔗
AI Orchestration Layer
Trigger · Validate · Analyze
  • Trigger Polls SharePoint on schedule (pull, weekdays)
  • Validate Checks file presence + format
  • On Fail Alert email — does not silently skip
  • Analysis 7-day rolling window + baseline comparison
  • Detection Spikes, dips, and threshold breaches
SP Poll → File Check → [fail] → Alert
File Check → [pass] → Read 7 → Baselines → Analyze
🧠
LLM Layer
Production Model
  • Model GPT Mini — won eval on quality/latency/cost
  • Evaluated GPT Mini/Nano, Claude Opus/Sonnet/Haiku, Gemini Flash/Lite
  • Prompts Iteratively evolved ("prompt skilling")
  • Few-shot What-good-looks-like examples embedded
  • Cost $0.50–$2.00 per daily run
🛡️
Grounding & Validation
Hallucination Fix: 40% → ≤5%
  • Problem v0 fabricated names, products, flagged wrong anomalies
  • Fix #1 Baseline file as source of truth for KPI targets
  • Fix #2 Iterative prompt refinement with few-shot examples
  • Fix #3 Field-level format validation at each step
  • Result ≤5% error ceiling — most runs zero errors
v0: 40% → + Baselines → + Prompt Skilling → + Field Validation → ≤5%
📧
Output & Delivery
Two Email Streams
  • Email #1 Executive Summary — strategic overview with trends
  • Email #2 Ops Alerts — action-oriented, flags KPIs to investigate
  • Timing Early morning, Mon–Fri
  • Impact Execs informed before workday; ops act proactively
⚠️
Error Handling
Failure Paths
  • Missing Alert email → workflow halts
  • Malformed Validation fails → alert → halt
  • API Down Step error → manual retry
  • No Fallback No automatic model-fallback (v2)
  • Quality Incorrect successful output not auto-detected (v2)
📊
Logging & Audit Trail
Split Architecture
  • System A Data Engineering — ETL and completeness logs
  • System B Orchestration platform — run/step logs
  • Tracked Pass/fail, step breakdown, cost per run
  • Gap No unified audit dashboard
🔄
Governance & Feedback
Human-in-the-Loop
  • Channels Email replies, Slack/Teams, ad-hoc messages
  • Used For Prompt refinements and baseline adjustments
  • Baselines Set with ops stakeholders, quarterly review
  • Gap No structured feedback form or scoring
🚧
Acknowledged Gaps — v2 Roadmap
Transparency on What's Next
1 Output validation before sending — no auto-check against source data
2 Prompt version control — no formal versioning or rollback
3 Schema evolution — new clients/KPIs may require manual updates
4 Recipient management — distribution list changes with roles
5 File retention — no archival or cleanup policy
6 Delivery confirmation — no verification against spam-filtering
🔧
Full Technology Stack
Production Infrastructure
GPT Mini (prod)Claude (eval'd)Gemini (eval'd)Relay.appSharePointAWS LambdaAWS EC2PythonSQLExcel (.xlsx)Email / SMTP