Flight Deck

I designed and built Flight Deck to reclaim the workday from email chaos. It’s an AI-powered triage station that didn’t just sort mail; it fundamentally changed how the team operated, turning a 10,000-email flood into a manageable stream of high-value conversations.

Book a call

Solution Architect & Tech Lead
Applied LLM/ML Engineer
Data & Integration Engineering
MLOps & Observability
Security & Compliance

Flight Deck auto-classifies incoming emails into prioritized queues.

~85%of inbound email auto-routed

~10,000emails processed per day

~3% misroute ratecaught by guardrails, never silently lost

Security and complianceGDPR compliant · Data stays in the EU · No training on client data, privacy by design · Encrypted in transit and at rest, audit-logged

The 10,000 Email Morning

Every morning, the team woke up to a disaster: 10,000 unread emails. Buried in the noise were urgent client requests, critical outages, and sales leads. But finding them meant sifting through thousands of "Thank you" replies, out-of-office notifications, and duplicate tickets. Morale was tanking, response times were slipping, and the "inbox zero" dream was dead.

Flight Deck architecture: ingestion and classification, a triage engine that auto-handles or routes to a human reviewer, then send, with a learning loop back into the model.

Taking Back Control

We built Flight Deck not just to filter email, but to understand it.

It starts with Ingestion & Classification: analyzing intent, sentiment, and urgency in real-time. But the real magic happens in the Triage Engine. It routes requests to the right human or handles them automatically, drafting context-aware replies that sound like us, not a robot.

Crucially, we kept humans in the loop. Flight Deck doesn't guess; it suggests. Agents approve or edit drafts, and every interaction teaches the model to be smarter tomorrow.

Flight Deck results: 90% noise reduction (10,000 daily emails to 1,000 meaningful tickets), first-reply time cut from hours to minutes, and urgent issues flagged in seconds.

From Chaos to Clarity

The impact was immediate and profound. We didn't just move numbers; we bought back time.

90% Noise Reduction: 10,000 daily emails became 1,000 meaningful tickets.
Instant Triage: Urgent issues were flagged in seconds, not hours.
Human Focus: The team stopped being data janitors and started being problem solvers.

The dashboard proved it: First-reply times plummeted, and customer satisfaction scores climbed. We turned a firehose into a fountain.

Stack & engineering notes

Python · Azure OpenAI (OpenAI-compatible APIs) · prompt versioning with staged rollouts · evaluation harness on curated test sets · guardrail layer for low-confidence routing · CI/CD · production monitoring and alerting · EU data residency, GDPR by design, no training on client data.

Three decisions carried this system. First, low-confidence classifications route to a human instead of guessing, which is why the misroute rate stays near 3% and nothing is silently lost. Second, prompts ship like code: versioned, evaluated against a fixed test set, and rolled out in stages, never edited live. Third, every automated action is logged and auditable, which is what made the system deployable under GDPR in the first place.

Project outcomes

• Strategic Calm: The panic of the morning backlog is gone. The team starts the day knowing Flight Deck has the watch.
• Consistent Quality: AI drafts ensure every customer gets a high-quality, on-brand response, every time.
• Scalable Trust: By showing our work (confidence scores, citations), we built a system that stakeholders trust to run the show.