Skip to content

Portfolio

What we've built

Real systems, built for real businesses. Client identities and production metrics are confidential — the architectures are not.

Where exact client figures cannot be disclosed, representative metrics are informed by published industry case studies and benchmarks, and are labelled as such.

Book Your Free Consultation See all capabilities

Case study

Bilingual WhatsApp agent for service businesses

Built on n8n, Python, Meta Business and the official WhatsApp Business API, powered by Claude. One multi-industry architecture serving cleaning, insurance and travel businesses — in English and Spanish.

  • Answers FAQs instantly
  • Qualifies leads
  • Captures structured info
  • Checks availability & books
  • Confirmations & reminders
  • Hands off to a human

The handoff is designed, not accidental: on low confidence, an explicit request for a person or a sensitive conversation, the agent stops and delivers a summary — contact, intent, history, suggested next action — so the customer never repeats themselves.

The architecture

  1. WhatsAppofficial Business API
  2. n8norchestration
  3. Claudeintent + language
  4. Actionanswer · qualify · book
  5. CRM + calendaralways in sync
  6. Humanhandoff with context

1,000+

conversations per month

Representative workload from published comparable WhatsApp deployments.

~65–70%

of routine interactions automated

Representative range from comparable published AI-agent deployments.

~3 weeks

to production

Representative timeline for a scoped WhatsApp deployment.

What the industry reports

Case study

An autonomous content system, in production

A Python daemon running 24/7 on a VPS. Nothing publishes without human approval — every piece passes a Telegram review first. Currently the content infrastructure behind two web properties, including the blog on this site.

38AI agents

  • Research
  • Trend detection
  • Drafting
  • SEO / GEO
  • Visuals
  • Distribution
  • Engagement
  • Analytics

8 specialised departments, one orchestrator.

The pipeline

  1. Trends
  2. Research
  3. Draft
  4. SEO / GEO
  5. Adaptper platform
  6. Human OKvia Telegram
  7. Publishsocial + blog
  8. Learnweekly loop

Cost governance

Model routing by cost and difficulty, daily budget with kill-switch. Internal technical cost: ~$0.10/post — model usage only, excluding human review, infrastructure and distribution.

Bilingual by design

EN/ES output with a voice profile per language, auto-maintained sitemap and blog index, and a hard "never invent statistics" grounding rule.

Live and verifiable

7+ articles published on this site's blog, both languages. ai-gentlab.com is the live demo.

Marketing OS

The productised version (in development): a modular skill engine with fail-closed QA, self-revision, credit billing and SEO/GEO skills for AI-engine citability.

What the industry reports

Third-party industry benchmarks — not measurements of this system.

Case study

A custom insurance CRM with AI where it counts

Built for an insurance agency's real sales process — HubSpot's methodology as inspiration, none of its rigidity. It replaced spreadsheets and inboxes with a single source of truth: leads, pipeline, tasks, follow-ups, scoring and history in one place.

  • Automatic lead scoring

    Source, intent, interaction, stage and recent activity — the team opens the day knowing who needs attention.

  • Context before every contact

    The full conversation history, summarised before each touchpoint. No scrolling through months of notes.

  • Follow-ups, drafted

    AI drafts and next-best-action recommendations. Where commercial judgement is needed, it proposes — never executes alone.

  • No lead left behind

    Leads without recent follow-up surface before they go cold, not after.

~10

team members supported

Representative deployment size.

500+

active lead records

Representative CRM workload.

What the industry reports

The FairBuild case is the published deployment behind the representative figures above.

Beyond the case studies

Everything we build

Organised by the four services behind them.

AI Agents →

Telegram agents

In production today: the approval bot behind our own content system. Human-in-the-loop from a chat, without touching a server or CMS.

Multichannel customer service

WhatsApp, web chat, email, voice and Telegram — EN/ES, with routing to a person by intent, risk or confidence.

Agent control centers

Fleets of specialised agents: one model per task, budget controls, automatic QA, human checkpoints, logging. Ours runs 38 agents in 8 departments.

n8n Workflows →

We don't sell "automations." We remove the manual steps between the tools your business already uses — the Portland Trail Blazers cut a 50h/week manual process to 3h ↗.

Integrations with your stack

HubSpot pattern: lead from form/WhatsApp/call → contact updated → AI classification + score → summary attached → task + owner → automatic follow-up, escalation past a threshold.

Content operations

The full pipeline from the case study above — research, drafting, QA, human approval, publishing, repurposing, analytics — as a service.

RAG Systems →

Your team shouldn't have to remember where the answer lives. At the far end of the scale, Nubank's internal knowledge tools reach 5,000+ employees monthly ↗.

What goes in

PDFs, SOPs, docs and FAQs → vector search, hybrid retrieval, reranking, permission-aware access — with cited answers, eval sets and guardrails.

Where it's used

Employee knowledge assistants, support copilots, policy lookup, SOPs, sales enablement, onboarding — standalone or wired into agents and CRMs.

AI Strategy →

Websites + SEO/GEO audits

This site is the live demo: bilingual architecture, structured data, internal linking, automated sitemap and GEO for AI-engine citability.

Operational dashboards

Sales, attribution, funnel, pipeline value, response times, automation health — one operational view instead of scattered spreadsheets.

Time-tracking for field teams

Clock-in/out, job assignment, timesheets, supervisor approval, payroll export. Built for teams tired of hours arriving over WhatsApp. Optional GPS, always with consent.

AI-first marketing operation

Our internal venture: research, content, SEO/GEO, distribution, lead capture, CRM and reporting — with human approval at critical checkpoints.

About the numbers on this page

Client identities and production metrics are confidential. Figures labelled "representative" are informed by published industry case studies and benchmarks from comparable deployments — they describe what architectures of this type achieve, not audited results of a specific confidential client. Third-party statistics link to their original sources. UI mockups are illustrative recreations, not real customer data.

Want something like this for your business?

Tell us what eats your team's hours. We'll tell you which of these architectures fits — including when none of them does.

Let's talk