Tanya C. Hallgren
AI implementation for agencies and in-house teams.
I take AI from experiment to working process.
- Focus
- Agentic AI systems and self-hosted models
- Currently
- Director of Development at Meticulosity, since 2021
- Open to
- Contract or full-time · Remote only
- Experience
- Twenty years shipping web projects
AI systems I have built and run
Five case studies: multi-agent operations, local retrieval with citations, human-in-the-loop pipelines, and AI on diagnosis, QA and scoping inside live client delivery. Each one says what it proves.
The agent team
In productionA multi-agent system on the Hermes framework, running a content operation in production since January 2026.
Two AI agents run a content operation's weekly duties on a schedule, with a person at every gate and a receipt for every run.
Proves agentic AI in production, eighteen scheduled duties running daily since January 2026, with the human gates that kept it safe.
HERMES · OPENCLAW · CLAUDE CODE · GROK · GPT · QWEN, LOCAL · PYTHON · SQLITE · FASTAPI · LAUNCHD · SLACK
RedPencil.ai
In the studioAn LLM analysis pipeline for fiction manuscripts, with local inference for the reading and a frontier model for the judgement.
An AI developmental editor for novelists, built alone and in public, phase by phase, with a library of what good books do filling first.
Proves a whole product designed, built and shipped alone, phase by phase, in public.
CLAUDE · QWEN, LOCAL · PYTHON · SQLITE · OMLX
Knowledge libraries
In productionLocal retrieval-augmented generation with citations, running offline on self-hosted models.
One ingestion engine and three separate libraries. Books go in on my own hardware, a person decides what is true, and the newest library answers questions with citations while offline.
Proves a local, citation-first retrieval system, 179 offline tests and 1,260 cited passages on its first day.
PYTHON · SQLITE FTS5 · OMLX · QWEN, LOCAL · GPT · CLAUDE · FASTAPI · SLACK
Team vault
In productionA shared-memory and coordination protocol for a human and AI agent team, enforced in git.
A shared, version-controlled memory for a team of one person and six AI agents across two machines, with the rules enforced at commit time rather than requested in a prompt.
Proves a coordination protocol for a mixed human and AI team, 457 hand-offs logged in its first ten weeks.
GIT · MARKDOWN · OBSIDIAN · PYTHON
Delivery at an agency
OngoingAI on diagnosis, QA and scoping inside a HubSpot Diamond partner agency's client delivery, and kept off the client code.
Browser agents for diagnosis and AI for QA, scoping and build plans inside live client delivery, with the review gates that let them survive a deadline. The client code stays hand-built.
Proves AI inside a real delivery schedule, kept to the work it belongs on, with the gates that survived it.
BROWSER-AGENT DIAGNOSIS · AI-ASSISTED QA · SCOPING AND BUILD PLANS · UNCHANGED REVIEW GATES · CLIENT DELIVERYAI implementation, agentic workflows, self-hosted inference
Five capabilities, each shown working in a case study aboveAgentic pipelines
Intake to vetted output: multi-agent workflows with a local model for volume and a frontier model for judgement.
Self-hosted inference
Local models on my own hardware, so client material never leaves the boundary it was given.
Human review design
The gates, diffs and override paths that decide what an agent is allowed to ship.
Delivery under deadline
Agents on diagnosis and QA inside live client work with a schedule attached, not a lab with no consequences.
Agents that build agents
The scheduled agents run skills and scripts I write with Claude Code. Build authority and runtime authority never sit in the same agent.
Ten phases, two built
One product, built alone, phase by phase. Bar height is the weight of the phase and brightness is its state. Choose any phase to read what it covers and what it taught.
The same file sets the status on the RedPencil case study.03 Foundation
IN THE STUDIOAuth, project shell, deploy pipeline, and the data model manuscripts hang off.
Current phase. The build plan and first design come from phase 02.
Building RedPencil.ai's foundation, the phase everything after it assumes: auth, the project shell, the deploy pipeline and the data model manuscripts hang off.
The two phases before it are done. The need was brainstormed and the product scoped, and the scoping had its own review loop: I scope with Claude, hand the scope to GPT to find the holes, then bring GPT's findings back to Claude to close them. Two models disagreeing is cheaper than one model agreeing with me, and the scope that came out of it is the one the phase map on the home page tracks.
The learnings library went from an idea to a running system in the same stretch, and it is the part of the project that works every night. A local model reads one published novel an evening and a frontier model records what the book does well, so the editor will have an answer key before it has a manuscript. One book a night is slow. It is also the pace the hardware allows without disturbing the agents that share it, so it proceeds, and the shelf grows.
Every plate here was made with generative AI — Mystic 2.5 Flexible, an older model I keep reaching for because nothing newer touches it for abstraction. I prompted them, culled them and chose the sequence; none was accepted as it first arrived.
A model proposes, a person decides what is good enough to keep. Knowing which model to reach for is most of the job.Start a conversation
AI implementation, contract or full-time, remote. Tell me about the role, or the problem behind it, and I will tell you honestly whether it is a fit.
