Systems we've designed and built.

How it works

Built on the OpenAI Agent SDK with a RAG (retrieval-augmented generation) layer. Company handbooks, benefits guides, and policy PDFs are chunked and embedded into Supabase's vector store. When a question arrives, the agent retrieves semantically relevant document fragments and generates a grounded response. Ambiguous queries trigger a structured escalation to an HR contact rather than a hallucinated answer.

How it works

The agent runs on the OpenAI Agent SDK with tool-use enabled for calendar lookups, door lock control, and maintenance ticket creation. Incoming WhatsApp and SMS messages are routed via Twilio. For each interaction, the agent determines intent, checks property context from Firebase, and responds in the guest's language. Only genuine complaints or unusual requests escalate to the host.

How it works

Built on Make with custom API calls to OpenAI for scoring and HubSpot for CRM updates. When a lead submits a form, the workflow enriches the record via a property data API, scores intent based on budget range, urgency signals, and enquiry type, then routes the lead to the most relevant agent based on availability and specialisation. High-intent leads also trigger an immediate SMS outreach sequence.

How it works

Built with LangChain and N8N. Documents uploaded to Google Drive trigger the pipeline. A Python extraction layer splits the contract into logical sections, passes each to a custom LangChain chain tuned for legal clause identification (indemnity, termination, IP ownership, liability caps), and runs a risk scoring pass. The final structured JSON is rendered into a formatted PDF report and emailed back to the requester.

How it works

Built on Zapier with a custom Python microservice for anomaly detection. Vehicle telemetry data (fuel usage, mileage, idle time) feeds into the pipeline on a schedule. The anomaly layer flags vehicles performing outside their baseline and categorises the deviation. OpenAI generates a plain-English summary for fleet managers. Reports are dispatched to relevant managers via email and Slack, with urgent anomalies triggering immediate alerts.

How it works

Custom GHL sub-account setup with pipelines mapped to the client's sales stages. Inbound leads from Google Ads and Meta are captured via custom funnels and automatically enrolled in AI-written nurture sequences segmented by service interest. Missed calls trigger an immediate SMS response. The booking flow connects to a confirmation automation that reduces no-shows through a 3-step reminder sequence. Dashboard reporting surfaces pipeline velocity and conversion by source.

How it works

Built with Google ADK for the agent orchestration layer. Inbound leads from multiple ad channels are enriched via data APIs, scored across 12 intent signals by an AI agent, and categorised as hot/warm/cold. Hot leads are enrolled in an immediate personalised outreach sequence. Warm leads enter a longer nurture track. Cold leads are suppressed. All decisions and reasoning are logged to the CRM for human review. The agent re-evaluates lead status on every new touchpoint.

How it works

Built on Next.js with a Python data layer and PostgreSQL for storage. A nightly ETL pipeline aggregates sales, ad spend, inventory, and returns data. An embedded AI layer (OpenAI API) generates natural-language commentary on trend changes, anomalies, and recommended actions — surfaced directly in the dashboard without requiring the team to write queries. Key views include product performance, cohort retention, and blended CAC by channel. Role-based access for operations, finance, and marketing teams.

How it works

Built from the ground up on Next.js (frontend), FastAPI (backend), and PostgreSQL with pgvector for semantic search. Clients upload documents via a branded portal. The processing layer uses LangChain to orchestrate extraction, classification, and risk-scoring chains. Results are stored in a structured database with full provenance tracking. The platform supports multi-tenant access, custom extraction templates per client, and export to Excel or a compliance-ready PDF. Built with SOC 2-aligned security architecture.

How it works

Python-based forecasting engine using time-series modelling against 24 months of sales history, seasonality signals, and marketing calendar inputs. When predicted stock falls below a dynamically calculated threshold, the system generates a purchase order, routes it for approval via Slack, and on approval dispatches it to the relevant supplier API. All decisions include an explainability summary. The Shopify integration keeps product availability status live across all sales channels.