Work / AI products
Streamed grant writer.
An LLM content generator that streams output over SSE as it writes, protected by a fail-closed prompt-injection guard and grounded against eligibility rules.
- Stack
- Gemini, Node.js, SSE
- Scale
- Production, multi-tenant
- Role
- Designed and built
Flow
- User brief
- Injection guard
- Constraint extraction
- Grounded prompt
- Gemini stream
- SSE to browser
- Draft
How it works
The user describes what they need in plain language. Before anything reaches a model, a guard classifies the input; if it looks like an injection attempt, the request fails closed rather than being sanitized and passed on.
A first pass extracts the hard constraints from the opportunity: word limits, required sections, eligibility language. Those become part of the grounded prompt so the draft cannot drift from the rules.
Generation streams token by token over Server-Sent Events, so the writer sees the draft appear as it is produced instead of waiting on a spinner. Length and structure are enforced as the stream arrives.
Hard parts
- Streaming through a load balancer and a multi-tenant API without buffering or dropped connections.
- Deciding what fail-closed means for a writing tool: refuse clearly, explain why, never half-answer.
Outcome
Became the most-used feature on the platform. Drafts arrive in seconds and stay inside the rules of the opportunity they are written for.