ericpendleton.com
Rev. Oct 2026 · Boston

G. Eric
Pendleton

Communications and AI enablement. Human judgment, governing AI.

Eric Pendleton, smiling, in a grey T-shirt against a white wall.

Training, quality systems and communications in pharma, healthcare and finance taught me to make complicated work clear, controlled and able to survive an auditor.

Now I bring that discipline to AI. I start with the problem, not the technology. I define the product, direct the AI that builds it, test it in real use, and keep at it until it behaves.

Contact LinkedIn ↗ Work samples About
Working systems

Two things I directed AI to build. Both are live. Go poke at them.

The Enterprise-D Flight Deck at red alert: shields up as a Borg Cube fires a green beam at the ship, with helm, tactical and long-range sensor panels around the view.

Take the conn.

Enterprise-D Flight Deck

A starship flight and tactical simulator. Scan anomalies, rescue derelicts, beam people around, blow things up. Fly the ship yourself, or let the crew do it for you, Captain Lazy.

37 days oldBuild 110Browser
Try it ↗
Origin story

Born September 10, 2026 · 37 days old

Why

I wanted to play an old Star Trek game. It doesn't exist anymore. The Windows version survives, but those things never run well on a Mac, and the juice wasn't worth the squeeze.

So I tried the online Star Trek games. I'm not a big gamer. The few I found were restrictive, stingy with free resources, and confusing. I just wanted to admire the Enterprise. Fly around. Fly it myself, or put it on autopilot and let the crew do the work.

On September 10, 2026, I decided to build my own.

How

The game became the experiment. Could someone who doesn't write code direct AI to build something truly complex, then test it, debug it from its own flight logs, and keep piling on features without losing control?

I set the direction and wrote the specs. I played every build and read the telemetry. ChatGPT and Codex wrote the code. Claude and Gemini reviewed each build independently before release. Saucer separation, fleet behavior, docking, mobile controls and far too many edge cases came out of that loop.

Turns out, yes. Build 110, day 37.

Source and build records on GitHub ↗

Unofficial, non-commercial fan project. Not affiliated with or endorsed by Paramount or CBS Studios.

LiveCue during its demo: a live transcript of a financing discussion with terms highlighted, and explanations of a term sheet, a nonrefundable deposit and a non-binding date in the side panel.

It's OK to zone out.

LiveCue

Live transcription for meetings that quietly explains the terms, acronyms and names flying past, and suggests a smart question to ask. No frantic Googling. No pretending you know what pari passu means.

v1.7.0-dev.22Web appAudio not storedTwo weeks, concept to reality

Works in Chrome or Edge, on a laptop or a phone (not Safari yet). On a phone, add it to your home screen and it opens like an app.

Try it ↗
How it was built

In complicated business meetings, financing, legal and tech terms flew by faster than anyone could look them up. So I built the help I wanted.

I defined the product and the interface, directed the AI that built it, tested it in real meetings, and had each release reviewed independently before it went out. Two weeks, concept to reality. The demo runs without a key. For live use, you bring your own Gemini key, which is never stored on the server.

Source on GitHub ↗

The Gaia dashboard, dark theme: a management update headline, a what-changed panel, a status bar and the first of three workstream cards, with figures, names and details blurred.

Gaia Development Dashboard Private

A complicated project shouldn't require a complicated explanation.

One view of financing, milestones, partners and dependencies for a live infrastructure venture. AI reads the week's documents, calls and correspondence and drafts the update; I check every figure against its source before it goes on the board. Built for facts that won't sit still. Figures and names blurred here. Demo on request.

The Hub: five platforms reading from and writing to one shared record, a sheet on a folder and a protocol, under the same rules. one record · same rules ChatGPT Claude Gemini Gmail Drive

The Hub In progress

Because “where the hell did I leave that?” is not a workflow.

I work across ChatGPT, Claude, Gemini, Gmail and Drive. Each one remembers its own conversations. None of them remembers the others'. So I built a boring source of truth with a smart layer on top: plain, static tools, a Drive folder, a sheet, a protocol document, that every AI reads from and writes to under the same rules. The static tools became the dynamic undergirding.

The payoff is that each project is about that project's goals, not about hunting down where the last decision went. The infrastructure is built. The write-up is slow. Going from frantic to organized is good for your blood pressure.

Garside's Ice Cream in Saco, Maine: a small white stand with a blue script sign and a line of customers at the serving windows on a sunny day.

The ice cream stand

Garside's Ice Cream, Saco, Maine

Garside's has made its own ice cream on Ferry Road since 1956. My family bought it the summer I finished high school: thirteen flavors, erratic hours, a loyal following. We fixed the hours, used better cream and added flavors, and revenue went up 800% in two seasons.

I made the ice cream and wrote the recipes, about a hundred of them, starting with Mint Chocolate Chip the first summer. Thirteen flavors became forty. I hired and trained every crew, set the prices each spring, placed the dairy orders, painted the building and mowed the grass, most of it on weekends from Boston. We sold it in 2018.

I didn't just sling ice cream. But I absolutely slung ice cream.

OperationsProduct developmentHiringTrainingEmployee manualPricingVendor negotiationsMarket research
First page of The Scoop, the Garside's employee manual

The Scoop

Employee manual · 10 pages · PDF · 0.9 MB

The training guide I wrote and kept rewriting. History, the "arc" scooping technique, opening and closing checklists, what to do about a Grumpster.

View ↗ Download
When the numbers slid, I studied it

The stand counts every order, every day, in a notebook. The count had gone from 62,000 orders a season to about 35,000. Instead of guessing why, I made Garside's the subject of my master's capstone: two surveys, 314 respondents; fourteen customer interviews and focus groups; a vendor-pricing review; state weather data; a season of watching the windows. A third survey followed in 2017.

  • Weather wasn't it. Four of the best ice cream summers in the record produced no bump.
  • The product wasn't it. 95% of customers rated Garside's equal to or better than Häagen-Dazs and Ben & Jerry's. 58% said better.
  • The decline tracked Maine's post-2001 tourism slump, shorter hours, and running out of basics like sugar cones. Once a customer is lost, getting them back is the hard part.
  • So the fix was consistency and promotion, not product. A fifteen-point plan, most of it cheap: new flavors every week (62% asked for them), a flavor contest (winner: Mosquito Ripple), weekday specials, gift certificates, the old hours restored, and a rewritten training manual.
  • The first move paid before the paper was finished. We asked the dairy for volume pricing. The base mix went from $26 a bag to $16, about $7,000 a season, and 41% of the target.

The uncomfortable finding: the stand was already doing most things right. The work was doing them every day.

Work samples

Two PDFs, an article and a cartoon walk into a bar.

The PDFs order a bourbon, neat. The article explains its manufacturing data. The cartoon picks up the tab.

First page of Selected Work Samples

Selected Work Samples

14 pages · PDF · 2 MB · August 2026

Eight samples from twenty years inside corporate L&D, plus the early AI record. A sole-authored white paper on competency-based training. A train-the-trainer course built and facilitated for a global biopharma client. A ten-step needs-analysis instrument. A reusable course template. A published article on AI text analytics. Onboarding guides for an enterprise NLP platform. Documentation at scale.

View ↗ Download
First page of the finance writing samples

Finance Emails and Blog Articles

4 pages · PDF · 0.8 MB · 2023

Four client pieces written for Pendleton Financing, my commercial-lending brokerage: why lenders want an LLC as the borrower, adaptive reuse, getting started in long-term rentals, and 1031 exchanges. Plain English for people about to sign something.

View ↗ Download

Luminoso Happy Halloween!

Short animation · YouTube

A basic Halloween cartoon I made for the Luminoso blog, years before the animation studio. Great fun. Best at 1.5× speed; the gear icon will get you there.

Watch on YouTube ↗

5 Steps to Unlocking the Value in Unstructured Big Data

Article · Quirk's Media · April 2017

Five practical things to do with the text nobody reads: survey comments, call transcripts, support tickets, social posts. Written for marketing researchers at a point when "text analytics" was a vendor promise more than a practice. Still holds up: the lessons are business strategy and operating habits, not the tools.

Read at Quirk's ↗
Demonstrations

I'd rather show you.

The examples are composites, built from the kinds of documents I've written and rewritten. None of them is a client's. The questions I ask before writing a word are the actual work.

The same facts have to land with four people who need different things from them. One version, four edits. The edits are where the judgment is.

The facts

SOP-0421 rev 7 goes effective Monday, November 3. The cleaning check moves to before batch-record sign-off. Retraining is required. Rev 6 is withdrawn the same day.

Asked first
  • Does the effective date fall before the last training session?
  • Can anyone still reach rev 6, including printed copies?
  • Does the system block execution without a training record, or only warn?
  • Which shift is exposed?

Operator on the floor

Starting Monday, November 3, the cleaning check in SOP-0421 changes. You'll record the swab result before you sign the batch record, not after. Your training is Thursday. Until you've completed it, you can't run the step. Rev 6 disappears from the system Monday, so don't print copies.

Date first. The change in one sentence. The gate as a consequence, not a rule. The last line anticipates the workaround.

QA manager

Rev 7 goes effective November 3. Training completion is the gate: no signed record, no execution under rev 7, and rev 6 is withdrawn the same day. The exposure is the night shift, which trains last; I've asked Scheduling to confirm coverage. Effectiveness check at 30 days against deviation rate on the step.

Names the exposure window. The shift that trains last is the risk. A measurable close.

Site head

One procedure changes November 3. No throughput impact if training lands by Thursday; a one-shift gap if it slips. I'd rather hold the effective date a week than run a shift uncovered. Your call by Friday.

Trade-off, recommendation, decision, deadline. Four sentences.

Auditor

SOP-0421 rev 7, effective 3 Nov 2026. Change: cleaning verification recorded prior to batch-record sign-off. Training records on file for all qualified operators prior to first execution. Rev 6 withdrawn 3 Nov. Effectiveness review 3 Dec.

Document numbers, dates, evidence chain. No adjectives.

How I work with AI

AI made me dramatically faster. It also gave me more places to lose the thread.

So every project runs on four rules borrowed from regulated work, where nothing is finished until someone else has checked it and the evidence is on file.

Spec before build

Nothing gets built until the problem, the audience and what "done" means are written down.

Independent review

A different model reviews the work. No AI grades its own homework.

Evidence before release

Tested in real use, checked against logs and sources. A claim without proof doesn't ship.

One current version

One record of the truth, updated in place. Not a pile of drafts that disagree with each other.

AI can generate learning content very fast. It cannot certify that anyone can now do the job.

Selected work samples, 2026

A tool will not fix a broken process any more than training fixes a hiring mistake.

Selected work samples, 2026

Generating information is easy. Keeping track of what actually happens, what matters, and what needs attention is not.

The Hub project sheet, 2026

In a 500-page knowledge base, a quiet fabrication survives review because nobody rereads page 340.

Selected work samples, 2026

The most common failure of AI training: a demo, a burst of enthusiasm, and no behavior change.

Selected work samples, 2026
About

Where the discipline comes from

I make complicated work clear enough to act on and controlled enough to survive an audit. That started in corporate training and quality systems at Takeda, Alexion, Shire, Citizens Bank and others, where nothing is finished until someone else has checked it and the evidence is on file.

Now I apply the same discipline to AI. At Lunchbox Labs, an animation studio where I'm a partner and executive producer, I run a human-governed, multi-model AI review workflow for contracts, operating agreements, financial records and investor materials. At Gaia Eco Developments, a regenerative-infrastructure venture where I'm an equity partner, I write the investor narratives, executive briefs and decision materials. The two projects above came out of the same habit: define the product, direct the AI, test it in real use, keep at it until it behaves.

Before that: six years in France, first teaching, then corporate communications and company-wide English training at bioMérieux in Lyon. Fluent French. Conversational Spanish. Boston, remote-first.

Currently
Partner and Executive Producer, Lunchbox Labs · Equity Partner, Gaia Eco Developments
Education
M.S., Business Management, Emmanuel College · M.A., French Language and Literature, Washington University in St. Louis, Teaching Fellowship · B.A., French Language and Literature, Colby College, Psychology minor
The full version

Born in Waterville, and raised in Saco, Maine. Only child. My father taught history and directed athletics at Deering High School in Portland.

The summer I finished high school, my family bought Garside's, a seasonal ice cream stand on the Saco River. I made the ice cream, wrote the recipes, hired and trained the crews, painted the building and mowed the grass. I kept doing that on weekends from Boston for years, to help my mother. She passed and I sold it in 2018.

Thornton Academy, Phillips Exeter, then Colby, where I studied French and sang tenor with the Colby 8. Then France: two years in Avignon running a school's term abroad, which I rebuilt from scratch; summers directing immersion programs for Dartmouth's Rassias program and Washington University; a stretch in Caen, Normandy.

Grad school in St. Louis, where I taught French at the Alliance Française and performed in rock musicals on the side. A biomedical-diagnostics company hired me to teach its engineers French, translate all kinds of business and technical materials, then moved me to its global headquarters in Lyon for corporate communications and company-wide English training. Six years in France all told. People still take me for French.

Boston since, specifically the South End. Software, banking, state government, pharma, in that order, mostly teaching people to use complicated systems and writing the documents that keep regulated work honest, and compliant. A few years running my own commercial-lending brokerage. Now a partner and executive producer at an animation studio with a feature in production, and an equity partner in an advanced infrastructure venture.

Also: a finalist on The Next MJ Millionaire, “Big Mike’s” business-competition reality show on Amazon Prime (Mike Straumietis, being the 6′8″ Founder and President of Advanced Nutrients). Fourth of sixteen, pitching ‘chill’ ice cream. SAG-AFTRA eligible. Jesus in Jesus Christ Superstar at Exeter. Formally practiced and always terrible at: tennis, piano, skiing, and windsurfing. Still a damn good singer. Maine accent available on request.

Contact

Available for fractional, contract and project work.

LinkedIn ↗ Boston · US Eastern · remote-first
↑ Top