Expertise / 03
LLM Systems
Models put to work where the answer has to be right: structured outputs, retrieval that must quote its source, and guardrails written in code, so the model drafts and deterministic rules decide.
11 systems, 6 delivered, 5 built
01 — Tools and stack
What this work is built on. Every tool below is part of at least one of the systems further down this page. HoverTap a card to see what it is used for.
OpenAI
Models
Classification, extraction and grading, each returning a structure that code can check.
Google Gemini
Models
Drafting and intent reading, from data that code has already confirmed.
Supabase
Retrieval
The store an assistant answers from: chunks, embeddings and documents in a database the client owns.
PostgreSQL
Retrieval
Similarity search with the permission filter inside the SQL, so a restricted document cannot leak through a prompt.
Google Drive
Retrieval
The source documents, checked every six hours. Only what has changed is embedded again.
AssemblyAI
Speech
Transcription into one timestamped transcript that every later step works from.
ElevenLabs
Speech
Voice agents, with everything they may not decide alone placed behind a tool they have to call.
n8n
Rules around the model
The gates: If, Switch and Filter nodes that decide whether anything happens next.
TypeScript
Rules around the model
Typed edit operations from a closed vocabulary. A change outside it is not something the model can express.
02 — Delivered
Lead-to-Cash CRM for Three Contractors
The model drafts and never sends. On the staff dashboard the rules-based score stays in charge, and the model writes its own column beside it.
AI Voice Agent System for a Training Academy
Eligibility is re-checked in the orchestration layer and overrides the agent unless the flag is explicitly true: the model proposes, a check decides.
Shopify Support Inbox, Handled End to End
Five agents, five structured output parsers: every model call returns a locked schema, so no branch depends on parsing free text.
Google Local Services Ads Lead Responder
Gemini drafts the reply and reads intent. Ten If nodes stand against four model calls, and nothing written in a prompt can book a job.
AdWash — Autonomous Google Ads Management
Campaigns pass through a generator, an adversarial critic, a deterministic strength gate and a grounding audit. AI edits are typed operations from a closed vocabulary.
Just Grade Metrics — AI Call Quality Assurance
Every score must quote the transcript. A validator confirms that the quote is there verbatim, and if it is not, the call is held back for a person.
03 — Built, not yet deployed
The workflow JSON is in the repository and imports into n8n; the node counts are taken from it. None of these has yet run against live accounts.
AI Front Desk — Missed Calls, After Hours and Booking
A price comes from a configured table, or the agent says it will not guess. Six triggers are matched in code against the running transcript, every turn.
Invoice & Document Intake into Your Accounting System
Extraction into a schema the model cannot bend, then arithmetic in code: line items plus tax must equal the stated total, within two cents, or nothing posts.
Company Knowledge Assistant That Cites Its Sources
The permission filter runs inside the search itself, and an answer must quote a passage that is a literal substring of what was retrieved, or it is refused.
The Weekly Number, Assembled and Sent
Every ratio is computed in code and shown with its working. The model that writes the narrative has no tools connected: it can phrase a number, not produce one.
Content Repurposing Line with a Human Approval Gate
The model picks only from windows already cut from the transcript, banned words are matched in code, and a person approves every draft.
04 — Book a call
Need a model you can trust in production?
Book a call to go through what the model should read and produce, and which decisions must never be left to it.