What problem does the Proposal Swarm (HITL) solve?

Sales agencies spend 5+ hours manually extracting discovery call notes to draft custom proposals, leading to acute pipeline bottlenecks. Standard automations and off-the-shelf LLM wrappers fail at creative deliverables because linear pipelines cannot handle ambiguity or hallucinate unfeasible commitments. Lack of strict data governance risked sending unverified pricing, incorrect timelines, or mismatched tech stacks to high-value enterprise prospects.

How does the Proposal Swarm (HITL) work?

Ingest & Parse Node: Accepts unformatted call transcripts, running speaker-diarization cleanup to isolate client pain points from casual dialogue. Scope Extractor Node: Leverages Claude 3.5 Sonnet with structured outputs to extract explicit technical requirements, implicit constraints, and timeline benchmarks. Cost & Timeline Estimator Node: Cross-references extracted requirements against a client-specific rate-card database in PostgreSQL, assigning baseline hour estimates and risk multipliers. Proposal Drafter Node: Synthesizes scope, timeline, and company case studies into an executive-ready Markdown document. Human-in-the-Loop Interrupt: LangGraph triggers an interrupt() before dispatch. Execution state freezes and persists to a PostgreSQL checkpointer. Next.js Review UI: The human reviewer inspects a side-by-side diff: raw transcript on the left, editable Rich Text Editor on the right with Approve/Regenerate triggers. PDF Generation & Dispatch: Upon approval, the graph resumes, compiles styled PDF assets via WeasyPrint, and queues client delivery.

Which tools and integrations does it use?

It is built on LangGraph · FastAPI · Next.js 15 and integrates LangGraph, FastAPI, Next.js 15, PostgreSQL, Claude 3.5 Sonnet, WeasyPrint, Docker, Supabase.

What results does it deliver?

Drafting time collapsed from 5+ hours to under 3 minutes. 100% human-verified accuracy with zero hallucinated pricing or deadlines. Sales close cycles compressed by 48 hours via instant turnaround. Unified pricing governance across all client proposals.

Autonomous Proposal Swarm with Human-in-the-Loop Governance

LangGraph multi-agent state machine with human-in-the-loop interrupts that turns raw discovery audio into executive-ready proposals in 3 minutes.