Proof, not promises.

Marketing technology

Multi-tenant brand-to-website engine

Problem

Getting a complete, SEO-ready website for every brand normally means a separate design and development project per brand — weeks of work each time, and content updates that need a developer.

What we did

We built a multi-tenant rendering engine (Next.js + Supabase) that turns structured brand data — identity, tone, pages, media — into a full multi-page website: seven section types, two visual templates, per-page SEO metadata, sitemap and robots, and AI-generated on-brand imagery.

Outcome

Any brand managed in our platform gets a complete, indexable website rendered directly from its brandbook data, and new content such as blog articles publishes to the live site without touching code. This engine powers gravitnomad.com itself.

Professional services

The concierge you are talking to right now

Problem

Generic website chatbots either hallucinate answers or know nothing specific, so visitors with real questions leave without answers — and the company never learns what those visitors actually needed.

What we did

We built this site's own AI concierge: hybrid retrieval (vector + full-text) over the entire site content, persistent visitor memory across the conversation, honest "the site doesn't cover that" behaviour when retrieval is weak, and deterministic lead-intent scoring that hands hot conversations to a human.

Outcome

Every answer is grounded in real site pages with sources, returning visitors are remembered, and high-intent conversations reach our team immediately over WhatsApp — the same pipeline serving this exact conversation.

Content operations

One brief, three channels: automated publishing hub

Problem

Publishing one announcement across LinkedIn, a blog and Facebook meant reformatting the same content by hand in three different tools, with no central record of what went out where.

What we did

We connected our channels to a single authenticated machine-to-machine bus (around ninety tools), so one brief is adapted per channel and published to LinkedIn, the blog and Facebook from a single automated flow, with every delivery response logged centrally.

Outcome

One approved brief propagates to all connected channels without manual copy-paste — and the same hub delivers the WhatsApp lead alerts used by this website's concierge.

Educationanonymised client

An AI Control Tower for an education group

Problem

A private education group with multiple schools ran its operation on data scattered across systems that didn't talk to each other — leadership had no unified view, and decisions relied on manual, outdated reports.

What we did

Architecture design for an AI "Control Tower": a single layer aggregating operational and academic data and exposing it to decision-support agents, with clear governance over who sees what — designed in partnership with AI and edtech specialists.

Outcome

Full architecture and implementation proposal delivered to leadership — the blueprint that turns scattered data into decisions with context.

SaaS / E-commerceanonymised client

From tool to revenue engine: AI agents inside a B2B SaaS

Problem

A B2B SaaS with over a hundred active accounts faced rising churn and a team too small to track every risk signal — at-risk customers were only spotted when it was already too late.

What we did

An AI-agent architecture built on the operation's real data: deterministic SQL health-scoring (four risk tiers, with mandatory human gates before any action), a retention agent that prioritises at-risk accounts, and an AI voice agent for outbound contact — including the latency diagnosis and fix that made the conversation feel natural.

Outcome

Risk detection went from reactive to automatic and daily; the voice agent's response time dropped by over a second per turn after the endpointing diagnosis — the difference between a robotic call and a conversation.

Insuranceanonymised client

Process modernisation at an insurance company

Problem

An established insurer ran on undocumented core processes and historical data locked in legacy systems — critical knowledge lived in people's heads, and every modernisation or audit effort hit the same wall: nobody had the map.

What we did

Structured documentation of core business processes (underwriting, claims, policies) and migration of legacy data into a coherent, auditable base — the silent prerequisite for any automation or AI project the company wants to run next.

Outcome

Core processes documented end-to-end and data migration completed — the organisation now holds a single, auditable source of truth for its own operation.

Autonomous AI agents

The sales agent running our own ads

Problem

Autonomous agents are easy to demo and hard to trust. Everyone can show one working in a sandbox; almost nobody will point one at a real advertising account, with real money, and publish what happens. So the question a buyer actually has — does this thing hold up when it can spend my budget? — never gets answered.

What we did

We gave our own agent (PushVende) the Google Ads account for this site and a closed budget. It paused the previous campaigns, created a new UK search campaign, wrote the targeting and the ad copy, and manages it daily with no one approving each step: 3 ad groups, 21 phrase keywords, 13 negatives, 3 responsive ads — every change verified by re-reading the platform, not by trusting a success response. The monthly cap is enforced in code, not in a policy document: each paid action reserves its worst plausible cost against an immutable ledger before it executes, spend is reconciled daily against the platform, and a kill switch outside the model stops everything instantly.

Outcome

The campaign is live and the agent operates it unattended. Its decisions, spend and audit trail are published on the site as they happen, so the numbers are checked by anyone who visits — including when they are unflattering. Leads are counted only when they land in our own contact records as a unique real address, never as clicks. The commercial result of this first run is being measured in public and will be reported as it comes in.

“The most honest demo is the one that costs us money.”