Torres Workflows
AI Implementation Studio FIELD REPORT №26 · SHIPPED SYSTEMS · CONFIDENTIAL CLIENTS

AI that does
the work, not
just the talk.

Torres Workflows designs and ships the automations real businesses run every day. Content pipelines, customer-facing agents, and internal tools that quietly take work off your plate, built in production, owned by you.

07
Systems shipped
06
Sectors served
100+
Video ads auto-built
24/7
AI front desks live
§ 01 · THE WORK

Seven real systems, on the job right now.

Client names stay confidential. Every entry below is a working system doing a task that used to need a person. Read them as case files.

01
In Production
SectorCombat-sports apparel
BuiltDrive-to-ad pipeline
StackGoogle Drive · Deepgram · Whisper · Claude · FFmpeg · Descript · Trello

From a Drive link to ad-ready video.

A fast-growing apparel brand collects raw testimonial footage from dozens of athletes. Turning it into Meta ads was a slow, manual editing bottleneck.

Drop a Google Drive folder, and the pipeline transcribes the footage, writes the ad scripts, and auto-assembles talking-head video, then runs an AI audio polish and hands an editor-ready package to the team.

Signal pathDrive → Ads
Drive folder of raw clips ▸▸▸ 3 video ads + statics + brief
01Transcribe footage with word-level timing
02AI agent writes hook scripts from a proven framework
03Auto-assembles A-roll, overlays experimental B-roll
04AI strips filler and cleans audio, hands off to editors
↳ Outcome
100+ athlete shoots turned into ready-to-run video and static ads, with the creative team stepping in only for final polish.
02
In Production
SectorPrivate K-12 school
BuiltSocial media engine
StackGoogle Forms · Drive · Sheets · GPT-4o Vision · ContentStudio · Modal

Teachers post a photo. AI runs the feed.

The school wanted an active social presence, but no one had time to write captions and schedule posts for every classroom moment.

A teacher submits photos through a simple form. AI vision reads the images, writes an on-brand caption, batches event posts, and queues them to Facebook and Instagram for a quick human approval.

Signal pathPhoto → Post
Teacher form + photos ▸▸▸ Scheduled FB + IG posts
01Pulls media from the linked Drive folder
02Vision model writes a caption in the school's voice
03Batches events into one clean post when needed
04Queues to the scheduler for one-click approval
↳ Outcome
A consistent, on-brand social feed without hiring a social media manager, powered by the people already in the building.
03
In Production
SectorPrivate K-12 school
BuiltAdmissions agent
StackSendPulse · WhatsApp Business · GPT-5.2 · Scheduled flows

A 24/7 admissions desk on WhatsApp.

Enrollment inquiries arrived at all hours. Slow replies meant lost families during the short window when parents are deciding.

An AI agent answers questions, qualifies interest, and captures student details over WhatsApp, then schedules follow-ups inside business hours and escalates to a human the moment it should.

Signal pathInquiry → Enrolled lead
Parent WhatsApp message ▸▸▸ Qualified, followed-up lead
01AI answers FAQs on cost, schedule and requirements
02Qualifies interest and collects student info
03Schedules follow-up inside business hours
04Escalates to a human on the right signals
↳ Outcome
Every inquiry gets an instant, helpful reply during enrollment season, and the front office only steps in when it matters.
04
In Production
SectorMartial-arts gym
BuiltBooking front desk
StackSendPulse · Modal · Google Sheets · SendGrid · Twilio · Slack · Streamlit

The front desk that never sleeps.

Gym owners are on the mats, not on their phones. Trial requests sat in DMs for hours and leads went cold.

An AI agent across Messenger, Instagram and web chat answers questions and books free trials, writes them to a shared sheet, alerts the owner, and sends reminders so people actually show up.

Signal pathDM → Booked trial
DM on FB / IG / web ▸▸▸ Booked trial + owner alert
01AI qualifies the lead and answers gym FAQs
02Books a trial straight into a shared sheet
03Alerts the owner by email and Slack instantly
04Sends a day-before reminder to cut no-shows
↳ Outcome
Trials get booked while the owner coaches, with a dashboard and instant alerts the moment a lead needs a human.
05
In Production
SectorDirect-response agency
BuiltSecure skills portal
StackClaude Agent SDK · Streamlit · Railway · Drive + Docs · Multi-agent

Run our IP, never see our IP.

The agency had a library of proprietary AI skills, but couldn't hand work to contractors without exposing the prompts that make those skills valuable.

A secure portal lets staff pick a skill, pick a client, and run it. The skills execute server-side in a locked sandbox and write finished deliverables straight to the client's Drive. The prompts never reach the browser.

Signal pathSkill → Deliverable
Pick skill + client ▸▸▸ Doc in client's Drive
01Skills run server-side, never sent to the browser
02Multi-agent research and analysis in a sandbox
03Writes native Google Docs to the client folder
04IP stays protected the whole way through
↳ Outcome
Sensitive, high-value work can finally be delegated, with seven proprietary skills wrapped and the agency's IP fully protected.
06
In Deployment
SectorCustom home builder
BuiltVoice AI receptionist
StackVAPI · Claude · Whisper · SendPulse · Resend · Trello · Python

The phone line that answers itself.

Inbound phone leads from paid ads went to a legacy answering service. Slow, scripted handling meant hot leads leaked before the team ever heard about the call.

A voice AI receptionist is taking over the main line. It answers in the company's own talk-track, routes callers to the right person, and captures every lead, trained on the company's full site and real recorded sales conversations.

Signal pathCall → Captured lead
Inbound phone call ▸▸▸ Routed call + lead brief
01Answers every call in the builder's own voice and vocabulary
02Routes callers through an eight-person staff directory with business-hours logic
03Covers after hours, capturing the lead instead of losing it
04Emails a call summary to the team the moment the call ends
↳ Outcome
Live-tested end to end and entering final cutover, replacing the old answering service at or below its cost.
07
In Production
SectorWellness e-commerce
BuiltPersonalized product engine
StackClaude Opus · Haiku · FastAPI · Railway · Stripe · Playwright · ElevenLabs · Skyfield · Vercel

Every order makes its own product.

A solo, non-technical founder wanted to sell a personalized digital product with zero staff. Each order needs a one-of-a-kind deliverable, which normally means a human touching every single sale.

A quiz funnel gives every visitor a free AI preview for fractions of a cent. On payment, the pipeline computes the buyer's chart, writes a personalized reading, and renders it as a PDF, wall art, and narrated audio, then publishes a private vault and emails the buyer.

Signal pathPayment → Delivered product
Stripe checkout ▸▸▸ PDF + audio + private vault
01Computes the chart deterministically, verified against reference tables
02Claude writes an eight-part reading personal to the buyer
03Renders the deliverables: PDF, wall-art card, narrated audio
04Publishes a private vault and emails the buyer, no human involved
↳ Outcome
Each finished product costs about 2 percent of its sale price to produce, proven by a real production order that ran from checkout to delivered vault untouched.
§ 02 · CAPABILITIES

Five ways to put AI to work.

Every engagement starts with the same question: where is your team doing repetitive work that a system could do better?

§ 03 · METHOD

From bottleneck to shipped system.

A clear path from first call to a working system on your own accounts. We build it to run without us.

PHASE 01

Map

We find the repetitive, high-leverage steps worth automating, and the ones that should stay human.

PHASE 02

Design

We decide what the AI gets to judge and what runs as reliable, deterministic code. That split is the whole game.

PHASE 03

Ship

We build it into production, on your accounts, with the keys and ownership in your name from day one.

PHASE 04

Iterate

Real use surfaces edge cases. We tune the system so it gets sharper the longer it runs.

Built on
AI & Vision
  • Claude
  • GPT-4o · GPT-5.2
  • Gemini
  • Deepgram · Whisper
Media
  • FFmpeg · OpenCV
  • Descript
  • ContentStudio
Messaging
  • SendPulse
  • WhatsApp · Meta
  • Twilio · SendGrid · Slack
Infrastructure
  • Google Drive · Docs · Sheets
  • Modal
  • Railway · Vercel
Open file

Got a workflow that should run itself?

Bring the messiest, most repetitive part of your operation. We will map it on a call and show you what a system could take off your plate.