Devin Huber
AI agents & workflow automation · n8n · Zapier · Claude

Agents that know when to hand it back.

I build AI agents that run real operations work: tagging content, reviewing copy for compliance, watching competitors, and writing weekly analytics briefs. Each one runs rules first, uses the LLM only where judgment is needed, and sends anything uncertain to a person. The ones that make decisions are scored against an answer key.

See the agents Book a live walkthrough
10agents & workflows built
in the past month
3companies given custom
spec builds
sample trace · metadata-enrichment
input
dedup
entity
llm tag
guard
AUTO_PUBLISH
EDITOR_REVIEW
SPECIALIST_REVIEW
Method

The same five stages, in every build

The tools change from project to project. The design stays the same, laid out here in the order data moves through it.

Stage 1 · Rules

Rules before reasoning

Regex, dedup and entity matching run first. The model only chooses among grounded candidates. For example, a common-noun check keeps "shell companies" from tagging Shell plc.

Stage 2 · Guard

Guards reject, not trust

Model output gets checked before anything moves. Codes that aren't in the vocabulary are rejected as HALLUCINATED_CODE. Quoted evidence that doesn't appear word for word in the source is rejected as UNGROUNDED_EVIDENCE.

Stage 3 · Route

Uncertainty goes to a person

Confidence floors are set per client and per category. If an LLM call fails, the item goes to review instead of being dropped.

Stage 4 · Output

Quiet by default

Briefs only raise an item when it's clearly needed. Scoring prompts check their own calibration: if most items score above 0.6 you are miscalibrated.

Stage 5 · Eval

The agent grades itself

Decision-making builds run against an answer key and report precision, recall, silent error rate, gating false positives and cost per 1,000 items.

Agents & workflows

Built this past month

Some are spec builds for hiring teams, some are live work for my current employer, and one is a personal project. Client names are withheld. The figures come from each build's own test runs.

n8n · 16 nodes
Built for
Legal & news information publisher (spec build)
Model
Claude Sonnet via HTTP
Stores
Data Tables (vocabulary)

Metadata Enrichment Agent

Adds controlled-vocabulary subject, industry, geography and company tags to incoming news as it arrives. It decides which articles can publish on their own and which need an editor or a specialist.

  • Near-duplicates are collapsed with 5-word shingles at Jaccard ≥ 0.50 before any tokens are spent.
  • The guard rejects made-up codes, wrong facets and evidence that isn't quoted word for word. Broader terms come from the taxonomy hierarchy, not the model.
  • Tested on a 24-article answer key with planted traps: "shell companies," passing mentions, retired-code bait and a renamed-company alias.
  • The run brief may only cite figures from the scorecard, and the Word export still ships if the brief call fails.
Tuning log, run 42: lowering the industry confidence floor to 0.70 raised straight-through from 20.8% → 45.8%, but silent errors went from 1 of 5 → 3 of 11. I reverted it and re-applied it only after completeness cues were added to catch under-tagging.
Open the evaluation scorecard ↗
37.5→16.7%error rate across 5 test runs
0.85subject confidence floor
100%company-match precision & recall
n8n · 8 nodes
Built for
Legal marketing agency (spec build)
Model
Claude Sonnet, batched ×5, 3 retries
Rules
ABA Model Rules 7.1 / 7.4, state variants

Bar-Compliance Content Reviewer

Checks law-firm marketing copy against attorney advertising rules before it goes live. Regex catches the obvious banned terms. The LLM catches what regex can't: implied guarantees, superiority claims and statistics without context.

  • Disclaimers are removed before scanning, because a compliant disclaimer itself contains the word "guarantee."
  • Each client has its own threshold (0.80–0.90) that separates blocking findings from advisory notes. Outcomes are BLOCK, REVIEW_REQUIRED, PASS_WITH_NOTES and PASS.
  • The scorecard shows which layer caught each violation. If the LLM never catches anything regex missed, it isn't earning its cost.
40ground-truth documents
10 · 5clients · state rule sets
7LLM violation codes
n8n · 3 workflows
Built for
Legal marketing agency (spec build)
Schedule
Monthly, per client
Stores
Snapshots + competitor registry

Competitor Watch & Monthly Digest

Monitors each client's competitor websites, detects what changed since last month, scores how much each change matters, and writes a digest an account lead can read in 90 seconds.

  • The registry builder finds case-results, attorney, awards and news pages based on what each page is, whatever the site calls it.
  • Change detection strips nav, footer and scripts, compares a djb2 hash, and passes only new sentences to the model.
  • A test fixture trims stored snapshots and backdates them 30 days, so the diff and scoring can be tested without waiting a month.
  • Scoring failures come back as REVIEW_MANUALLY, never dropped. The run reports a signal ratio to catch a threshold set too low.
0.65materiality threshold
5 · 7calibrated bands · categories
90starget read time
Zapier Agent
Built for
Leadership-development institute (current work)
Sources
GA4 (2 properties) · Kajabi
Schedule
Every Monday

Monday Operations Brief

An analytics agent that emails leadership each week with the biggest changes, green and red flags, and anything worth optimizing across site traffic and the course platform.

  • Reads as an informative email, not a to-do list. It offers suggestions and never says "we have to."
  • On quiet weeks the suggestions section is left out entirely, instead of printing "nothing this week."
  • When items do come up, they're ordered by ease and importance together, with no effort or impact labels.
  • Tested against the live account, which surfaced a connector that ignored its tag filter and a report action limited to one date range. Each run has a 10-step budget and a restricted tool list.
Rev 4built from live-account testing
10step budget per run
0filler sections on quiet weeks
Zapier · 2 agents
Built for
Leadership-development institute's podcast
Record
Google Sheets
Output
Scores, decline drafts, monthly digest

Guest Application Flywheel

Two agents that score incoming podcast guest applications, draft personal decline emails for weak fits, and send leadership a monthly digest of the pipeline.

  • Safety comes from limiting tools: each agent holds only the actions its job needs.
  • Application text is treated as data. The agents are built to ignore instructions hidden inside applications.
  • Writes happen in a set order, so the sheet never shows a half-processed row. A person can overrule any score.
15point scoring rubric
5scored criteria
2agents with separate tool sets
n8n · form trigger
Built for
AI note-taking product for social care (take-home)
Model
Claude Sonnet, temp 0.1
Companion
Pipeline & quality dashboard

Assessment Template Test Harness

Runs any conversation transcript through an assessment template, so prompt changes can be regression-tested on real input instead of eyeballed.

  • Never infers a diagnosis, capacity, risk or eligibility, and never softens a risk someone stated.
  • Contradictions aren't resolved. Both sides are recorded with who said them, and figures are kept exactly as spoken.
  • The companion dashboard runs 15 statistical tests with multiple-comparison correction, recalculated in the browser from raw rows.
Open the companion dashboard ↗
0.1temperature
11self-check items
15corrected stat tests
n8n · 17 nodes
Built for
A family member's house search (personal)
Schedule
Daily 8am · Sunday top 10
Model
None, fully rule-based

House Search Daily & Weekly Digest

Turns noisy listing-alert emails from three sites into a star-rated daily digest, then sends a Sunday email with the week's best ten and an interactive board.

  • No LLM. Listing text is scored with weighted condition phrases ("new roof" +3.0, "as-is" −6.0) scaled to 0–12.
  • A 15-term auction and foreclosure filter logs the phrase that caused each listing to be dropped.
  • Dedup runs in three passes: listing ID, then normalized address plus ZIP, then across sources in the same batch.
  • Missing fields score at the midpoint and are flagged on the card, not silently penalized.
126point weighted score
13scoring factors
3-passdeduplication
swipe, drag or use ← →
HTML builds

Dashboards, reports & build sheets

Hand-built, self-contained pages that go with the agents: evaluation reports, interactive analysis and the specs the agents are built from. Pages that contain client data are shown live on a call.

Metadata Agent Scorecard

Live ↗

Evaluation report for the enrichment agent: results against the editors' answer key, findings from each test run, and a 90-day rollout plan.

  • precision / recall / F1
  • cost per 1,000
  • rollout plan

Notes Pipeline Insights

Live ↗

Interactive dashboard analyzing an AI note-taking product's pipeline and summary quality, filterable by team, template and period.

  • 4 tabs
  • funnel + hours-saved slider
  • 15 stat tests

House Search Board

On request

Interactive map and ranked list of 239 listings, each with a star rating and a score breakdown you can expand.

  • SVG map
  • 11-factor score
  • filters + detail panel

Signal to Pipeline

On request

Diagnosis of a training company's marketing and sales funnel from CRM, LMS and survey data, with seven proposed automations ranked.

  • drop-off analysis
  • cohort revenue
  • phased roadmap

Monday Brief Build Sheet

On request

Build spec for the analytics agent, with a log of fixes found by testing against the live account and a copy-paste instruction block.

  • fix log
  • step budget
  • tool restrictions

Guest Flywheel Build Sheet

On request

Two-agent spec covering the scoring rubric, which tools each agent gets, defenses against instructions hidden in applications, and a test plan.

  • 15-pt rubric
  • injection defense
  • test plan

House Digest Runbook

On request

Setup guide for the listing digest: inbox labels and filters, the parser, the dedup table, and why it uses alert emails instead of scraping.

  • runbook
  • Gmail filters
  • dedup ledger
Toolkit

What I build with

n8nworkflows · Data Tables · Code nodes ZapierAgents · MCP Claude APISonnet JavaScriptin-flow logic HTTP / REST Google Analytics 4 Kajabi Gmail Google Sheets HTML · SVGdashboards
About

Operations first, then automation

I'm an AI Solutions Analyst in Columbus, Ohio, building and evaluating AI agents for a leadership-development institute.

Most of my work starts with a person doing something by hand every week: pulling numbers, reviewing copy, screening applications. I work out what they check, what they'd never let slide, and when they'd want to be asked. Then I build an agent that follows the same rules, and I measure its accuracy before anyone relies on it.

My degree is in finance (B.S.B.A., The Ohio State University, 2025), so I frame automation in terms of business impact: hours saved, errors caught, cost per item.

View résumé (PDF) LinkedIn ↗

Strengths

  • Multi-layer agent design
  • Hallucination & conformance guards
  • Statistical testing
  • Human-in-the-loop routing
  • Evaluation harnesses
  • Accuracy & false-positive measurement
  • Prompt engineering
  • Webhook & API integrations
  • GA4 & KPI reporting
  • Excel modeling
  • CRM workflow design
  • Requirements & process mapping
  • SOPs & enablement
Contact

Want to see one run?

I'm happy to screenshare any of these builds: open the workflow, run the test set, and walk through the scorecard, including the runs that didn't work.

d.huber.professional@gmail.com LinkedIn ↗