John Hill

Analytics and AI for telcos that need decisions to stick.

Manager of Communications and Ecommerce·GVTC Communications·San Antonio metro

§ 01

Now

Building an LTV:CAC model on the GVTC subscriber base. Shipping NPS sentiment classification with telco-specific labels. Standing up Atlas, an internal AI agent with Slack-native delivery.

§ 02

Selected work

  1. 01

    Telco Churn MLOps Pipeline

    Built a public-safe churn pipeline that turns synthetic broadband customer data into validated scores, action routing, API predictions, monitoring reports, dashboard views, and passing CI.

    Churn models usually die as one-time spreadsheets. The useful system is the repeatable path around the model: validation, leakage checks, training, scoring, serving, monitoring, and business routing.

    A standalone MLOps portfolio project that scores 12,000 synthetic broadband customers, separates retention candidates from winback/address-lifecycle opportunities, and exposes the workflow through FastAPI and Streamlit.

    • Python
    • LightGBM
    • FastAPI
    • Streamlit
    • GitHub Actions
    Case study GitHub

  2. 02

    Customer Growth Phase II Model

    Well-received by leadership April 2026. Phase 2 expansion in flight with rental capture rate, county ownership verification, and offer-hook lift testing.

    GVTC's residential growth was stalling under fiber overbuild competition with no household-level penetration model.

    A household-resolution model that maps every rooftop passing to its active-subscriber status, layered with offer-hook lift testing across rental and ownership segments.

    • SQL Server
    • Python
    • Claude Code

  3. 03

    Atlas — internal AI agent

    Took over routine analytical workload via a multi-agent AI system with scheduled EC2 crons, Slack-native delivery, and isolated session workspaces per sub-agent specialty.

    Routine analytical work and cross-system orchestration was eating practitioner time, and a single AI agent couldn't cover the breadth required without thrashing.

    A Slack-native AI agent built on OpenClaw with custom MCP integrations, a sub-agent layer for domain specialization, identity files, and isolated cron sessions. Dual-auth across Anthropic subscription and API spillover keeps it always-on.

    • OpenClaw
    • Anthropic API
    • EC2
    • MCP
    • Slack

  4. 04

    Custom Claude Code skill graph

    Turned Claude Code into a domain-aware ISP analytics workbench through 115 custom skills covering data exploration, GVTC vocabulary, daily ops, content workflows, and AI debugging.

    Off-the-shelf AI tooling didn't know GVTC's data model, ARPU conventions, vendor systems, or the cooperative's voice.

    A composable skill graph with reference files for telco domain, iVUE/NISC schema, voice-DNA, and single-purpose workflows from morning briefings to NPS sentiment classification.

    • Claude Code
    • Anthropic API
    • MCP
    • Python

  5. 05

    vault-search MCP server

    Shipped a local MCP server that serves semantic search over a 30K-vector personal knowledge base into Claude Code workflows.

    Knowledge captured across hundreds of vault notes was inaccessible during AI workflows without manual lookups, breaking the find-it-then-think-with-it loop.

    A Python MCP server bridging Obsidian Smart Connections embeddings (bge-micro-v2, 384-dim) to the Claude Code MCP protocol with sub-100ms query latency.

    • Python
    • MCP
    • Obsidian
    • sentence-transformers

  6. 06

    NPS Sentiment Pipeline

    Surfaced billing and reliability themes the manual coding pass had missed across 18 months of comments.

    Open-ended NPS responses were summarized by hand and key themes were getting lost.

    A telco-specific theme classifier with billing, speed, support, reliability, and pricing labels.

    • Anthropic API
    • Python
    • SQL Server

  7. 07

    Meyer Ranch / Code Red

    Pivoted the retention metric from new-customer MRR to net-change-plus-churn-avoidance, which became the competitive response template for fiber overbuild defense.

    High-penetration subdivisions were getting overbuilt by Spectrum fiber and the existing retention math didn't capture the value of churn avoidance.

    Modeled subdivision-level penetration and retention against incumbent and new-build scenarios.

    • SQL Server
    • Python
§ 03

How I think

  1. Verify inherited claims before propagating them.
  2. Most analyses are decisions in disguise. Surface the decision early.
  3. Bring the join-key reality, not the ideal-world ER diagram.
  4. ARPU is recurring, not billed. Patronage credit is a member rebate.
  5. Use AI to lower the cost of asking the next question. Save your judgment for the answer.
  6. If the audience already has a report for this data, match its labels. Continuity beats parallel terminology.
§ 04

Contact

LinkedIn
linkedin.com/in/johnhill387
Substack
practicalmachines.substack.com
GitHub
github.com/jhillgvtc
Email
commsjhill@gmail.com