Stackpine

Systems that pay for themselves.

Selected systems in production use: autonomous agents, on-prem LLM inference, a live SaaS platform and internal AI tooling. Each one replaced hours of manual work with something that runs itself.

Stackpine illustration: software windows and a data network set among pine trees

Selected builds

Email Triage Agent

Autonomous inbox management

50 emails triaged in 13s, down from 8 minutes

A persistent agent that sorts high-volume Outlook inboxes automatically and rescues legitimate mail from junk.

Full detail

The problem

High-volume Outlook inboxes needed constant manual sorting, and legitimate contacts occasionally landed in junk unseen.

What was built

Persistent Python agent that authenticates to Microsoft Graph (MSAL device-code flow, cached token), applies sender rules first, batch-classifies the rest with an LLM (Actionable, Automated, Spam), moves messages to the right folders, and rescues protected contacts from junk. Replicated across multiple staff inboxes from a reusable build prompt.

Impact

Cut triage of 50 emails from roughly 8 minutes to about 13 seconds.

Stack

  • Python
  • Microsoft Graph API
  • MSAL
  • Anthropic API
  • Local JSON classification cache
  • Continuous polling service

On-Prem AI Server

NVIDIA DGX Spark

Nemotron Super 49B running fully on-prem

An on-premises inference layer that keeps compliance-sensitive AI workloads entirely in the building.

Full detail

The problem

Compliance-sensitive workflows (intake, case-adjacent processing) cannot run on cloud LLMs.

What was built

On-premises inference layer serving a large local model through an OpenAI-compatible endpoint, powering local document RAG, privacy-sensitive processing, and scheduled agent runs as persistent services.

Stack

  • NVIDIA DGX Spark
  • Ollama
  • Nemotron Super 49B
  • systemd services

RevLane / OpenThrottle

Revenue recovery SaaS

Production SaaS · verified by 56 tests

A SaaS platform that recovers declined-service revenue for auto and powersports shops.

Full detail

The problem

Auto and powersports shops lose revenue when customers decline recommended service or walk out on a quote, and nobody follows up.

What was built

SaaS platform that tracks declined services and walk-out leads and runs automated SMS follow-up to recover the revenue, with shop dashboards, admin tooling (impersonation, health-check watchdog, daily digests, trial and churn automation), and a performance-based pricing engine. Verified by a 56-test production suite.

Stack

  • Node/Express
  • React
  • SQLite
  • Railway
  • Vercel
  • Telnyx 10DLC with Ed25519 webhook signature verification
  • Stripe
  • Resend

Staff Knowledge Assistant

Internal RAG chatbot

~30 hrs/week of interruptions absorbed

An internal AI assistant that answers staff policy and procedure questions from the organization's own documents.

Full detail

The problem

Staff interrupted supervisors constantly for policy and procedure answers buried in documents.

What was built

Internal AI assistant grounded in ingested organizational documents (PDF, DOCX, TXT) with OCR for scanned files, streaming answers, voice input, a feedback loop, per-document re-indexing, and honest-decline behavior when documents do not cover a question. Upgraded from document lookup to an operational reasoning advisor.

Impact

Absorbs an estimated 30 hrs/week of procedural questions and supervisor interruptions.

Stack

  • Node/Express
  • React/Vite PWA
  • SQLite FTS5
  • tesseract.js + pdfjs-dist OCR
  • SSE streaming
  • JWT auth
  • Railway
  • Automated daily Box backups

Resume Screening Pipeline

Privacy-first candidate triage

Zero PII leaves the building

A privacy-first pipeline that scores resumes against a job description with a locally hosted LLM.

Full detail

The problem

A large resume backlog plus a live stream of new applicants, with strict data privacy requirements.

What was built

Pulls resumes from Google Drive, extracts text, scores each candidate against the job description on weighted dimensions (JD match, behavioral adaptability, AI/tech leverage) using a locally hosted LLM so no PII leaves the building, writes ranked results with rationale to Google Sheets, logs failures for review, and covers both a one-time backlog burn and an always-on watch folder. Framed as triage assist with human-in-the-loop review, never a hiring decision-maker.

Stack

  • n8n
  • Google Drive
  • Google Sheets
  • Ollama on on-prem hardware

Operational Dashboards

Role-specific live views

~15 hrs/week saved across roles

Role-specific dashboards that replace manual status compilation across Outlook, Smartsheet, and internal trackers.

Full detail

The problem

Leadership and program staff compiled status manually across Outlook, Smartsheet, and internal trackers.

What was built

Role-specific dashboards with live task, fleet, grant, and communications views, plus email notifications sent through Microsoft Graph.

Impact

Estimated 15 hrs/week saved across roles.

Stack

  • React
  • Node/Express
  • SQLite
  • Microsoft Graph API
  • Smartsheet API
  • Tailwind

Funeral Triage

Dispatch coordination

In active paid use by a commercial client

A dispatch coordination app built for a direct-to-families funeral service, running in production today.

Full detail

The problem

A direct-to-families funeral service needed structured dispatch coordination in place of ad-hoc phone and text handoffs.

What was built

A dispatch application that organises incoming cases and coordinates assignment, built for and paid for by the client and in active daily use.

Want systems like these?

Tell us what’s slowing you down. Free quote, plain English, zero pressure.

Get a free quote See client results