Business Plan Generator
A multi-agent system built with LangGraph that autonomously researches markets, analyzes competitors, and generates comprehensive, structured business plans. Agents communicate via state graphs to refine financial projections and marketing strategies.
The Problem
Generic 'AI business plan' tools tend to produce plausible-sounding numbers that aren't tied to anywhere real — a rent estimate that's really just a language model's prior, not what a space actually costs in that neighborhood. I wanted a planner that grounds its financials in live local data and is honest when its own budget assumptions turn out to be wrong.
Approach
Built a LangGraph state machine with a sequential spine (parse location → build a 3-tier blueprint → pick a tier against budget) that fans out into five research agents running concurrently — market pricing, competitors, rent, permits, payroll — then fans back in before compiling the plan.
Gave each research agent a grounded web-search tool via Groq's native browser_search on gpt-oss-20b for live, cited lookups, with an automatic fallback to DuckDuckGo (and a cooldown once a rate limit is hit) so one flaky search doesn't take down the whole run.
Added a budget-gate node that re-checks the researched total (equipment + legal fees) against the user's actual budget after real numbers come in, rather than trusting the LLM's pre-research capex guess — if it's still over budget, the graph downgrades a tier and loops back through research automatically.
Pulled real nearby competitors from OpenStreetMap's Overpass API, trying multiple public mirrors in sequence since the main instance drops connections under load.
Replaced a fixed-timer progress bar with a real one: a streaming endpoint emits a Server-Sent Event the moment each LangGraph node actually finishes, so the UI reflects genuine agent progress — including a visible 'downgrading tier' state when the budget-gate loop fires — instead of a countdown that had no relationship to what the backend was doing.
Key Features
- State-machine driven agent orchestration using LangGraph.
- Market Research Agent scraping real-time web data.
- Financial Projection Agent generating 3-year P&L forecasts.
- Outputs a formatted 15-page business plan document.
Challenges
The financial logic was the trickiest part to get right: an early version locked in feasibility using the blueprint agent's rough capex guess, made before any research had run, which meant a plan could be labeled 'feasible' off a number nobody had actually verified. Recomputing feasibility twice — once as a fast pre-research gate, once for real after the researched equipment and legal costs were in — and looping back to a cheaper tier when the real numbers didn't fit was what made the budget math trustworthy. Streaming was its own problem: with a single blocking endpoint, the frontend had no way to know which agent was actually running, so the progress UI was pure guesswork. Switching to LangGraph's `astream(stream_mode='updates')` over Server-Sent Events fixed that, but it meant designing the UI around genuine parallelism — five research agents finish in whatever order their searches happen to resolve, not neatly one after another, so the progress display had to represent that as a group rather than a fake sequence.
Outcome
A planner that researches instead of guesses: real competitor names from OSM, live rent and material pricing pulled via search, and a feasibility verdict that's checked against actual researched costs rather than an LLM's opening estimate. The progress UI now shows exactly what the agent graph is doing in real time, and a health-check ping smooths over cold starts on the free-tier host so a slow first load reads as 'waking up' instead of a silent hang.