Joseph Haenel

A delivery-support application for a fictional distributor, with evidence and human review built into the workflow.

OpsDesk checks customer-account permissions before retrieving policies and shows citations alongside order and delivery records. Employees review and edit saved proposal revisions before approving the exact action.

Approval and simulated case creation commit in a single PostgreSQL transaction. Stable operation IDs, duplicate prevention and saved results support safe retries and recovery.

The current demo uses template drafts, with no live model call. All data and cases are synthetic; customer responses are never sent. Developed with AI assistance.

FastAPI · React / TypeScript · PostgreSQL · Policy retrieval · Human approval · Docker

An enterprise AI platform used by more than 1,000 employees since January 2025.

I built durable chat execution with queued jobs, idempotent requests and reconnect recovery. Permission-aware SharePoint retrieval uses delegated Microsoft Graph access, resumable indexing and atomic activation of completed indexes.

Release workflows include database migration checks, controlled traffic promotion and rollback. I added OpenTelemetry instrumentation for model and tool execution, including a Monte Carlo evaluation, to support debugging and cost visibility.

I also built assistant feedback analytics: clustering unanswered questions, tracking review status and providing location-filtered dashboards to help business teams identify gaps in assistant responses.

A self-service assistant-builder pilot enabled 10 employees to create 15 custom assistants, with approval, version history and sharing controls.

Azure AI · Python · PostgreSQL / pgvector · Redis · Microsoft Graph · OpenTelemetry

A knowledge-capture platform with evidence and human review built into the workflow.

I built and deployed the application using FastAPI, React and hybrid retrieval. Evidence-linked extraction, human approval and durable background jobs support the capture process, with Okta authentication for access.

FastAPI · React · Hybrid retrieval · Durable jobs · Okta OIDC

Forecasting workflows designed to be evaluated, reproduced and operated reliably.

I own warehouse forecasting in Python and Snowflake. The work includes reproducible backtests, shadow evaluations, incomplete-data checks and baseline fallbacks for planning feeds.

Python · Snowflake · Backtesting · Model evaluation · Data quality

An internal career-matching application refined through employee testing.

I built job-data ingestion and embedding-based matching, then adjusted the weighting of education and work experience in response to tester feedback. The application work included session management, backend and interface integration, and PDF career plans.

Later work brought Career Coach into DotChat and added a dashboard, with a redesign that keeps submitted career information local to the request.

Python · Embeddings · Recommendation systems · Application integration

AI-assisted project intake with human review before work enters Jira.

I built an authenticated application for refining stakeholder requests and preparing structured handoffs. Reviewers approve Jira Epic and Story creation, with context for downstream coding agents included in the handoff.

Next.js · PostgreSQL · Okta OIDC · Jira integration · Human approval