Inspiration
Solo founders can move quickly with AI, but turning a raw idea into a coherent plan is still daunting. HardLaunch serves as an AI operator that captures the founder’s concept, preserves context across sessions, and applies the kind of planning rigor you would expect from an experienced advisor - so decisions are grounded in evidence rather than guesswork.
What it does
HardLaunch runs an agentic workflow that converts ideas into actionable strategy. A conversational intake captures the problem, audience, solution, moat, and constraints, while a context manager maintains a live summary of the venture that persists across interactions. Specialized planning, finance, and market agents then generate focused analyses and recommendations. A retrieval-augmented generation layer draws from uploaded research to surface evidence and citations. All results appear in a navigable founder dashboard where you refine strategy, inspect sources, and seamlessly hand off tasks to the appropriate agent.
How we built it
The system is implemented with Google’s ADK agent framework, using Gemini 2.5 Flash for reasoning and models/embedding-001 for vector search. Documents are stored in a LlamaIndex vector database persisted under src/data/persist, with Whisper/YouTube ingestion hooks to handle rich media. A FastAPI backend exposes an /api/chat endpoint for the UI, and the ADK session service shares state so onboarding, context management, and tool calls remain synchronized.
Challenges we ran into
Reliability was the central challenge. General-purpose LLMs often fill in gaps with plausible guesses, which is unacceptable for business planning and financial decisions. We addressed this by enforcing retrieval-first prompting, surfacing citations for claims, and allowing agents to defer when evidence is missing. The result is a workflow that favors accuracy and transparency over speculation.
Accomplishments that we're proud of
We’re proud of a multi-agent architecture with clear role separation and shared context, which reduces drift and improves coordination. By anchoring analyses to a curated knowledge base through RAG, the system consistently produces evidence-backed outputs. The “live brief” that updates as founders answer questions or upload research keeps every agent aligned and makes progress visible in real time.
What we learned
Agent quality is a direct function of prompt design, persistent memory, and clean, relevant data. Guardrails that let the system acknowledge uncertainty lead to better outcomes than confident but unfounded advice. Exposing citations, assumptions, and next steps inside the interface helps founders make decisions faster and with more confidence.
What's next for HardLaunch
Next steps include expanding ingestion to Notion, Google Drive, and CRM exports, and strengthening the RAG toolkit with inline citation tracking and source quality scoring. We plan to add automated progress reports and milestone tracking so plans evolve into living roadmaps that agents can monitor and update. We will also enable auditable, autonomous function calls to market data APIs to keep trend analyses current and easily visualized.
Log in or sign up for Devpost to join the conversation.