Grants AI

Overview

Grants AI manages the full grant lifecycle in one place. It creates grants, organises documents, researches past winners, drafts applications, generates reports with live programme metrics and tracks each email, deadline and missing award letter automatically. It serves nonprofits and grant teams who run all this across folders, spreadsheets and inboxes today. They lose time and opportunities to the admin work around applications rather than the applications themselves.

Challenges

The existing grants implementation could not be safely shipped. It carried a backlog of pre-merge blockers around permissions, tenant schema context, AI stubs and broken client imports. Access control gaps and incorrect schema handling created a real risk that one organisation's grant data could show up in another's. A vector-extension database migration sat in the wrong place and threatened to disrupt other tenants sharing the same cluster. On the AI side research could silently accept ungrounded past-winner results. That quietly fills a grant record with plausible but unsupported information. Arguably that's worse than returning nothing. External research pulled in URLs with no validation which exposed the system to local files and internal network targets. Drafting was one-off and disposable. There were no persisted draft records, no status tracking and no way to regenerate a single weak answer without redoing the whole application. Reporting had no safe path to live programme metrics and risked accidental writes back into upstream modules. Grant lifecycle work was untracked. No email classification, no deadline alerts, no auto-archiving and no detection of missing award letters. And valuable funder knowledge locked inside recorded webinars was never captured.

Solution highlights

We remediated all pre-merge blockers first. That meant permissions, tenant schema context, AI stubs, the client import path and vector migration placement, plus removing an unnecessary PDF dependency. Then we built the capabilities properly. Past-winner extraction now uses structured AI output so research fills database fields reliably. Grounding verification hard-fails on unsupported results rather than silently downgrading them. URL validation rejects local files, metadata IP ranges and internal network addresses. Drafting became a persisted trackable process with a draft model, a status state machine and an AI drafting service using retrieval and prompt assembly. Per-question regeneration means a single weak answer can be rewritten without touching the rest. Reporting gained a grant report model with funder templates and a cross-plugin metrics interface. That interface reads live programme data with no write access to upstream modules. Grant lifecycle email ingestion runs through a dedicated mailbox with an email classifier, auto-archive rules, missing-award-letter flagging, deadline notification, in-app alerts and an email digest. It all shows up as a correspondence timeline on the grant page. Video and webinar ingestion generates transcripts then chunks and embeds them. Funder guidance recorded in a webinar becomes searchable context during drafting and reporting.

Key Features

• Pre-merge blocker remediation across permissions, tenant context, AI stubs and migrations
• Past-winner research using structured AI output that populates database fields
• Grounding verification that rejects unsupported research instead of downgrading silently
• URL validation blocking local files, metadata ranges and internal network targets
• Grant application draft model with status state machine and persisted records
• AI drafting service with retrieval, prompt assembly and per-question regeneration
• Grant report model with funder-specific templates
• Read-only cross-plugin metrics interface substituting live programme data into reports
• Grant lifecycle email ingestion with classification and auto-archive rules
• Deadline notifier, missing-award-letter flag, in-app alerts and email digest
• Correspondence timeline on each grant page
• Video and webinar ingestion with transcripts, chunking, embedding and retrieval

Our Role & Approach

Our role was to make an ambitious AI feature set trustworthy enough to rely on for funding applications. A confident-sounding error has real consequences here. That shaped two decisions in particular. First we made grounding verification hard-fail. If a research result can't be supported it is rejected rather than quietly accepted with lower confidence. Unsupported data in a grant record is worse than a gap. Second we limited the reporting layer to read-only access to upstream metrics. Reporting can never accidentally write into the modules that own that data. Security work came before feature work. Tenant isolation, permission enforcement and the misplaced vector migration were all resolved before anything new was built on top. We also time-boxed the riskiest research component which was automated web browsing. A structured-text fallback was defined in advance in case it proved impractical.

Technology Stack

Backend: Django plugin, Celery tasks, pgvector, AI vendor orchestration. Frontend: React, TypeScript. Integrations: Gmail API, YouTube, yt-dlp, Whisper/OpenAI API, Playwright, Gemini grounding. Security: Gmail OAuth, Google service account, schema context enforcement, granular grant permissions.

Outcomes

Grant teams can ship the platform with confidence because the pre-merge blockers and tenant-safety risks are resolved. Tenant data stays isolated with corrected schema handling and vector migrations moved onto a cluster-safe path. Past-winner research lands as reliable structured data rather than unverified free text. Each result is grounded and verifiable and unsupported outputs fail fast instead of slipping into an application. External research is safer by design and unsafe or internal URLs get rejected before they can be followed. Drafters work from persisted records with status tracking. They can regenerate any single answer without redoing the whole application. Drafting isn't a one-shot gamble any more. It's an iterative process. Reports pull live programme metrics through a read-only interface which keeps upstream modules protected. Grant lifecycle management is automated so correspondence, deadlines, alerts, missing award letters and archiving are tracked without manual chasing. And funder knowledge from videos and webinars becomes searchable context that strengthens drafting and reporting.