Atlas
Agent creation by conversation.
Overview
Atlas is an internal blueprint automation platform built to remove the manual overhead of configuring AI agents for each new client engagement. Before Atlas, every new support, sales, or analyst agent meant hand-writing prompts, filling JSON configs, and repeating the same setup steps. Atlas replaces that with three creation paths: a conversational AI builder that asks the right questions and auto-generates the agent configuration; a structured form for operators who know exactly what they want; and a PDF upload pipeline that parses specification documents, cleans the extracted text with GPT, and populates the blueprint automatically. A FastAPI backend handles prompt generation via OpenAI, multi-step blueprint assembly for each agent type, file processing, and the full authentication and deployment flow into the underlying agent runtime. Internal teams can go from a client brief to a deployed, live agent without writing a prompt or touching the deployment platform directly.
How it works
Operator selects: guided chat builder, structured form, or PDF specification upload.
Chat answers stored in session; form fields validated; PDF extracted via pdfplumber then GPT-cleaned.
GPT-4o-mini generates a persona, objectives, and conversation script tuned to the agent type.
Full agent config assembled; analyst agents require sequential API calls per extraction target.
Blueprint linked to the target organisation if not already; team context substituted into the config.
Blueprint deployed to the agent runtime; operators see real-time step-by-step status in the UI.
Results
What was built
Three creation modes — conversational AI builder (guided chat, session-stored answers, streaming generation), structured form (field validation before generation), and PDF upload (pdfplumber extraction → GPT cleaning → auto-populated blueprint). All three produce the same blueprint schema.
OpenAI prompt generation pipeline — GPT-4o-mini takes the user's raw answers and generates a persona, a set of objectives, and a conversation script tuned to the agent type. Runs with configurable retries to handle TPM rate limits under load.
Multi-step analyst blueprint builder — analyst agents are assembled across multiple API calls: step one creates the agent, subsequent steps create individual extraction targets (display name, description, criteria, data type, restrictions). Each step references the previous response via a key-chain pattern.
PDF specification parsing — pdfplumber extracts text from uploaded client spec documents; a GPT pass cleans garbled OCR output; the backend then maps the cleaned content to persona, objectives, and script fields. Supports CSV and Excel specs via pandas for analyst extraction targets.
Token and session management — three concurrent token scopes (super admin, user, deployment) with automatic refresh and fallback-to-login logic. In-memory session store (UUID-keyed Map) tracks conversational builder state across multiple frontend steps without database round-trips.
End-to-end deployment flow — blueprint created locally, assigned to the target organisation if not already linked, then deployed to the agent runtime with team context substitution. The UI surfaces status at each step; operators can also browse, filter, and delete blueprints by org.
Role
Built the full-stack platform on a small internal team — FastAPI backend, Next.js frontend, prompt generation pipeline, PDF parsing, multi-step blueprint logic for three agent types (support, sales, analyst), and the end-to-end deployment integration.
Stack