Oathkeeper - An agent that answers "Should we take this AI dev deal sales brought in?" with a Go/No-Go decision
Translated by AIView original
Screenshots
Screenshot 1 of 5
B2B AI dev shops usually decide whether to take a project through verbal reports and meetings, so everyone uses different standards and it can take days to reach a conclusion.
Risks only show up after you've signed. I built this to turn that process into an evaluation with consistent criteria.
Pick a deal from Notion, attach a memo or document (Word/PDF) if you need to, and start the analysis.
- Deal structuring: extracts client, amount, deadline and technical requirements from free text like meeting notes
- Scoring: rates the deal out of 100 on 7 criteria, including technical fit and profitability, then returns Go / Conditional Go / No-Go
- Resource estimate: calculates staffing, duration and cost based on our in-house headcount rates
- Risk analysis: identifies risk factors across technical, schedule, financial, client and scope
- Similar cases: finds the 3 most similar past projects
- Report: builds an executive summary and saves it to Notion
Deals that lack key information don't get forced into a score. They're sent back as 'On hold'. Evaluation weights and agent prompts can be edited right from the admin screen, and the logs show what the agent received and how it judged at each step.
How it was made
Built it with Claude Code, from planning all the way to deployment settings.
-
Documentation Used a meta prompt to write a PRD, split the PRD into a Phase 0~9 checklist (task-plan.md), and had it implement one phase at a time. CLAUDE.md holds the evaluation criteria, decision thresholds and commands, so context carries over even when the session resets.
-
Agent design Built with LangGraph in this order: deal structuring → resource estimate and similar case search (parallel) → scoring and risk analysis (parallel) → final decision. Profitability scoring needs the calculated cost first, so I split the parallel stage into two steps.
-
Prompt separation Prompts are pulled out of the code into YAML (Jinja2) templates, so you can tune them from the screen without redeploying.
-
Testing In a notebook I checked each piece one at a time: LLM → prompt → node → graph. Then added tests with pytest and Playwright. I also used Context7, shadcn, Playwright MCP and a code review sub-agent.
Stack: FastAPI, LangGraph, LiteLLM (GPT-4o, Claude Sonnet), Pinecone, PostgreSQL, Next.js, shadcn/ui, Docker Compose
- Built with
Comments 0
No comments yet
Be the first to comment.