VishnuJogappagari

A CV arrives by webhook. A parser agent extracts it into a schema. A fraud detection agent looks for fabricated history. A bias detection agent strips the signals that shouldn't influence a decision. A scoring agent ranks against the role, and an interview generator writes the questions. Each of those runs its own Gemini model behind a structured output parser, so the next node gets typed fields, never a paragraph to re-parse.

Around that spine sit three more pipelines: a recruiter copilot that answers plain-language candidate searches, a rediscovery pass that re-scores past applicants against new roles, and a hyper-search ranking engine over the whole Supabase pool. A global error handler logs anything that falls out.

Triggered from Drive, Gmail or a schedule. The document is extracted, then twenty-two code nodes carry the redaction and rule logic before anything reaches a model. The secure edition exists because law firms cannot post client identity into an API. Five model passes handle clause classification and red-flag detection, and the result lands in an Airtable register with Telegram and Gmail escalation for anything critical.

On top of it sits an operator console with a WebGL sector background, a risk distribution ring, a live activity feed, and a decision row: approve, escalate to counsel, or reject.

Form submissions and inbound email merge into one path. Attachments are extracted from PDF, an agent classifies the matter type, and a confidence gate decides what happens next. Above the threshold it proceeds: Drive folders are created for the case, a consultation is booked on calendar, the client gets a confirmation, the relevant department gets a brief, and Sheets logs the row. Below it, the matter is flagged for review and a holding response goes out instead.

The canvas is organised into labelled zones (trigger, core processing, case handling, onboarding, low-confidence reminders, error handling) because someone other than me has to be able to read it.

The front is a Next.js site with a booking flow and a voice showcase, built around one line: your business misses calls, ours never do. Behind it is the part that matters: a thirty-one node backend wiring the client dashboard to Supabase, to Dograh for the agent runtime, and to a Vobiz voice stack for telephony.

A prompt builder compiles agent instructions through Groq and pushes them to Dograh. A separate deployment recipe makes the whole thing reproducible on a Coolify VPS, with a verify command that has to exit clean before a client is handed the keys. A red verify means not delivered, whatever the dashboard says.

A Gmail trigger catches the question, retrieval runs over the firm's indexed filings and correspondence, and the reply is composed only from what came back — then sent in-thread and logged. There is no path where the model answers from memory, which is the entire point when the question is about a specific client's specific filing.

It doesn't run alone. A document intake agent keeps the index fed from inbound email and Drive, and a compliance deadline monitor watches the calendar side, so the corpus the answers come from stays current instead of decaying.

The dispatcher runs on a schedule, rotates five Gmail accounts so no single mailbox carries the volume, and writes a unique opening line for every lead with a model before prepending it to the template. Three follow-up stages run off the same sheet.

In parallel, five IMAP listeners sit on those inboxes around the clock. A reply halts that lead's sequence immediately and writes the reply text back to the sheet, so nobody who has already answered ever gets a follow-up. That's the failure mode this engine was built to remove.