Stephen Minemann

Automated pipeline that replaced a manual multi-hour weekly analyst workflow for tracking aging incident tickets (open > 30 days) across field ops teams. Ingests the weekly ticket export, classifies every ticket via a layered rule + TF-IDF fallback classifier (fully offline, no external model dependencies), infers actionability from blocker signals in ticket text, and rolls results up across region → area → market → technician. Tracks week-over-week change via an append-only snapshot so the impact of each weekly nudge is measurable over time. Delivers an interactive web dashboard with three views — days-open summary matrix, issue-theme breakdown, and a technician drilldown with a color heat scale — plus a one-click per-person email export with full ticket detail attached as a spreadsheet. All consumers read from a single structured JSON payload, decoupling the transform layer from the presentation so the delivery can move from static files to a live REST endpoint with no front-end rework.

// internal project · actively in development

Built the first repeatable outage-time dataset for the Nokia Radio Replacement program after a director-backed PMO request revealed no method existed to connect project schedules with national alarm data. Translated Nokia schedules, site lists, and maintenance windows into Snowflake SQL parameters to isolate qualifying alarms by site, project, market, and time window. Defined downtime logic and transformed raw alarm extracts in pandas into filterable CSV/Power BI outputs distributed through SharePoint for PM validation and leadership review. Productized into a reusable Snowflake aggregation table once cross-market demand grew.

// internal project · code and values have been sanitized and made ambiguous

Matching features combining normalized address tokens, geospatial distance, fuzzy similarity, and candidate ranking to link cell-site inventory to premise IDs. Designed a validation and exception-tracking schema for match status, review outcomes, metadata gaps, and audit history.

// internal project · code and values have been sanitized and made ambiguous

Repeatable cleanup workflow for site-compliance data: reconciled 3,425 missing-site records and 7,046 MWR compliance rows across Snowflake source exports, field assignment files, and site metadata. Standardized raw compliance values into pass/fail/remediation buckets and generated owner-ready exception queues. Identified 1,013 one-field correction candidates packaged as correction-ready outputs.

// internal project · code and values have been sanitized and made ambiguous

Production-grade Python CLI that reads CSV/XLSX site lists, authenticates to power-cabinet web interfaces, collects battery state-of-charge and remaining runtime, and exports outage triage reports sorted lowest battery to highest. Env-based credentials, timeout controls, HTTPS to HTTP fallback, masked IP logging, debug snapshots, and timestamped outputs. Built to run safely during disaster recovery.

Site-level Cost-Per-Site model for tower crew support and material spend across 6,000+ cell sites. This gave leadership a repeatable metric instead of one-off cost pulls. Normalized expense categories, surfaced outlier sites, trend views, and priority filters. Helped reduce average CPS from ~$180 to ~$150, beating the $160 target.

// internal project · code and values have been sanitized and made ambiguous

Built an end-to-end personal health data pipeline that ingests Apple Watch and iPhone biometrics, stores them in an analytical database, and serves aggregated metrics and time-series trends through a REST API. Handles 55 metric types across a full year of data. Containerized with Docker and deployed to AWS EC2, with DuckDB persisted to S3 so storage survives compute restarts. Ported the summary endpoint to AWS Lambda to contrast always-on vs. serverless execution models. Provisioned the full environment — EC2 instance, S3 bucket, IAM roles, and security groups — as code using Terraform. Key engineering problems solved: inconsistent field names across Apple Health metric types, DuckDB concurrency under async requests, duplicate accumulation on re-import, and S3 IAM permission behavior that returns 403 instead of 404 for missing objects without explicit ListBucket rights.

Private Telegram bot that runs a local LLM via Ollama on a home Linux server — no cloud API required. Supports multi-turn conversation with per-user history, automatic web search injection for time-sensitive queries (news, weather, scores), daily morning briefings with LLM-summarized headlines, and flexible one-off reminders. Authorized to a single Telegram user ID, deployed as a systemd service that auto-starts on boot and restarts on failure.