Agentic CV Builder

An agent that interviews you about your background and produces a one-page PDF. Rather than generating a document from a single prompt, it decides what it still needs, asks one specific question at a time, and saves each section as it learns it — so the conversation and the CV stay in step. The preview updates as you talk, and what you download is a real PDF.

The design constraint is that the model never authors content freely: it fills a schema. Every tool call is validated with Pydantic before it can change any state, and a malformed payload comes back to the model as an error it has to correct rather than failing the turn. It is forbidden from inventing anything — employers, dates, degrees, skills — and refuses when asked to embellish. Presentation is not its job either; colours, fonts, density and layout are yours, and the agent has no tool that can touch them.

The PDF is drawn with ReportLab from that same document, in either a two-column design or a single-column layout that automated parsers read more reliably. The live preview is a server-side rasterisation of the actual PDF rather than a separate HTML mock-up, so the preview cannot drift from the file you download. Smaller decisions target machine readers as much as human ones: bullets are drawn as text glyphs instead of vector circles, because a circle disappears from extracted text and collapses a job’s achievements into one run-on paragraph — which is what an applicant tracking system, or an LLM screening your CV, actually reads.

The renderer also measures how much space is left on the page, so the CV can report when it no longer fits. Sessions live in Redis with a 24-hour expiry and are never written to a database. Token counts are shown in the interface because cost per session was a design input rather than an afterthought: document state is summarised instead of resent in full each turn, the tool loop is bounded, and styling never calls the model at all — which keeps a complete CV in the region of a cent.