About / Roadmap — Funding Monitor

future (planned)

The items below were originally sketched as two separate near-term milestones ("0.4" and "1.0 — commercial platform") before 0.5's infrastructure work shipped ahead of them. Folded into one backlog now that there's no fixed next version number attached — these are the acknowledged next priorities, not dated commitments.
GOAL (commercial platform): turn this from a single-tenant tool into a real multi-tenant product — sized against ~200 CONCURRENT users. This is a foundational rearchitecture (auth + tenancy touch nearly every route), not an incremental feature. Nothing in this section is built yet.
Authentication: real login (currently none — the live site has no auth wall at all, anyone who finds the URL can reach every page and every action)
A role-BUILDER, not just fixed tiers: Platform Admins define roles by ticking which specific tools/features that role can access (checkboxes against real permissions — "manage sources," "override matches," "manage billing," "invite users," "access AI provider settings," "run bulk actions," etc.), rather than being stuck with a hardcoded 3-tier list. Ships with a couple of sensible default role templates (Platform Admin / Organization Admin / Organization User) so most organizations never need to touch the builder, but anyone with unusual needs (e.g. a read-only auditor role) isn't stuck
Consultancy model: a single account that manages MULTIPLE organizations (e.g. a consultancy handling several client profiles) — needs a many-to-many user↔organization relationship, not the current implicit "whoever's running the app owns every profile" model, plus a way to switch/scope which organization's data is in view
Data model changes this implies: a User table, a Role/Permission table (see role builder above), an organization-membership table (many-to-many), and every existing query (Opportunities, Matches, Sources, Overrides) scoped by the requesting user's accessible organization(s) and permissions instead of returning everything unconditionally
Admin-configurable settings instead of hardcoded constants — starting with the matching-weight blend (AI_SCORE_WEIGHT/TAG_OVERLAP_WEIGHT in matching/engine.py, currently 65/35 fixed in code) and extending to things like candidates-per-check defaults, taxonomy edits, and which sources are shared vs. per-tenant
Commercial-readiness: billing/subscription handling, usage limits per plan, and deciding what's shared infrastructure (sources, the taxonomy, AI provider config) vs. what's genuinely per-tenant (organization profiles, matches, overrides, workflow state)
IMPORTANT sizing point: AI compute does NOT scale with USER count — Opportunity rows are already global/shared (extracted once, used by every tenant) and Match rows are keyed per (opportunity, organization profile), not per user. 200 concurrent users spread across a modest number of organizations is a far smaller AI workload than 200 separate organizations would be — cost/throughput planning should size against expected ORGANIZATION count, not user count
A real cloud AI provider (Claude/OpenAI) as a proper fallback, not just a nice-to-have — the live site currently depends on the office's own Ollama being reachable over the internet (static IP + port-forward + firewall rule), which is a genuinely fragile chain (office internet, power, or IP changes would silently break AI extraction on the live site)
Monitoring/alerting stack — nothing beyond manual server-admin-GUI checks exists yet; no automated uptime/error alerting
Finish the off-VPS backup story — the server-side pull mechanism (restricted, chrooted, read-only SSH account) is ready, but the actual scheduled pull job on the user's own backup device still needs configuring on that device
Security hardening beyond the deployment basics already in place: rate limiting on login and on AI-triggered/expensive endpoints (stops one runaway client or bug from generating a large, unexpected AI bill), an audit log for role/permission/billing changes and match overrides, and per-organization AI usage/cost tracking (useful for spotting anomalies early and for usage-based pricing later)
A staging environment to rehearse risky changes (schema migrations, auth rollout) somewhere reversible before they touch the live database
Partner recommendations from an Odoo contact export (companies + tags) matched against each call's domain/technology tags — needs a decision on live sync vs. CSV export and a tag-mapping scheme between Odoo's tags and this app's taxonomy
Document/template link extraction for non-eu_portal (HTML/RSS) sources — currently eu_portal-only, since only its conditions text is parsed with links preserved
Per-source candidate-link tuning for PRIMA/Interreg (news-heavy sites without a clean listing→detail-page structure produce a lot of AI calls on non-call pages)
Scheduled unattended periodic source checks
Claude/OpenAI providers exercised end-to-end alongside Ollama
Cross-source duplicate detection (the same call sometimes appears on more than one portal)
Email/notification alerts on new high-match opportunities
Proposal-drafting assist
Calendar view of upcoming deadlines

v0.5 (current)

Moved from a local-only tool to a live, containerized deployment: https://funding.ecect.eu — real HTTPS (Let's Encrypt), deployed onto shared self-hosted infrastructure (Docker + Nginx Proxy Manager) alongside another project, reusing its existing reverse-proxy/TLS tooling rather than building a new stack
Database moved from SQLite to PostgreSQL — all real production data (opportunities, matches, sources, organization profiles, overrides, settings) migrated and verified row-for-row; local SQLite + the original in-process threading remain fully supported for local development, unaffected
Background jobs moved to a real priority queue (Redis + RQ) in the deployed environment: three queues ("high"/"default"/"low"), always drained in that order — one-off user-triggered actions (add-by-URL, PIF upload, recompute matches, test source) get priority over routine bulk work (the full EU-portal sync, the multi-model benchmark sweep). Replaces the old "any job blocks any other job equally" behavior
Automated nightly database backups with an actually-tested restore (not just an assumption it would work) — plus a dedicated, tightly restricted pull mechanism (chrooted, read-only, key-auth-only SSH account) so backups can be pulled off the server on a real schedule
The office's local Ollama instance is now securely reachable by the live deployment (scoped firewall rule + router forwarding), so AI extraction works on the hosted site using existing local hardware rather than requiring a cloud AI subscription on day one
Two real cross-database bugs found and fixed during the move to PostgreSQL: a SQLite-only performance tweak (`PRAGMA journal_mode=WAL`) that was firing on every database connection regardless of which database was in use, and a startup race between the web process and the background-job worker both trying to set up the database schema at once

v0.3 (shipped)

EU Funding & Tenders Portal: fixed a pagination bug that silently capped discovery at 500 of ~750+ live calls (missing exactly the near-term-deadline ones); results now sort soonest-deadline-first instead of newest-start-date-first
EU portal opportunities now carry real structured fields straight from the portal's own API (not AI-guessed): Programme, Call identifier, Type of action, Deadline model, Planned opening date, Model Grant Agreement type, and Submission status (Open/Forthcoming) — plus a "Load more details" button to backfill older rows
Fixed a bug where a candidate's identifier was re-derived from its URL instead of carried from discovery, which silently failed on ~78% of EU-portal candidates in one observed run; also fixed the source's "last checked" timestamp permanently showing "never" whenever any candidate in a run failed
Deadline-range filter, results-per-page control (100/250/500/1000), and a configurable per-source "candidates processed per Check now run" cap (was a fixed 60) so a full catch-up run can process everything in one go
Domain (~30 EU-portal topic categories) and Technology (Cascade Funding Hub's own taxonomy) tagging — AI-classified on every opportunity, human-picked (colour-coded chip picker) on organization profiles — plus a Field-of-expertise picker for profiles. Both feed into filtering and into the AI match-scoring prompt
Matching-prompt rewrite so domain/technology tag overlap (including a profile's more specific tag matching an opportunity's broader one) is explicitly weighed in the score and named in the rationale
Fixed two real hard-filter bugs found via user-made overrides: eligibility phrased as plain "EU countries" (not spelled out as "European Union"/"member state") and "EU entities" (a blanket org-type allowance, not a specific type) were both silently excluding genuinely-eligible organizations — affected 74 of 720 existing match rows
Human-in-the-loop match override: correct any match's score by hand with a reason, without ever losing the system's own original score/rationale; a dedicated /overrides/ page reviews every correction across the whole app and flags rule-filter corrections specifically as likely code bugs; AI-scored matches also feed a profile's past corrections back into future scoring prompts as calibration
Star/favorite flag, a "Starred only" filter and "Starred first" sort
Workflow tracking per opportunity — New → Viewed (automatic) → In progress → Decided (manual) — with a filter row and a quick "In progress" link
Best Match column/sort: defaults to best-match-first (was deadline-first), and now shows every organization's score, not just the top one
Document/template links (application form, evaluation form, MGA, guidance annexes) extracted from the EU portal's own conditions text and shown on the opportunity detail page — not yet available for non-eu_portal sources (see 0.4)
Light/dark theme toggle, defaulting to light regardless of OS preference (previously dark-preference systems had no way to switch within the app)
Add an organization profile from an uploaded Partner Identification Form (PDF or Word) as an alternative to crawling a website — same AI drafting and human-in-the-loop review as a URL crawl, including auto-tagged domains/technologies/expertise. The full extracted document text is kept permanently on the profile as a reference, viewable on the review screen
Update an EXISTING profile from a newly-uploaded PIF (not just create a new one) — re-drafts every field in place and always drops back to "pending review" so the redraft gets a fresh look before it's trusted for matching again
"Add organization" consolidated to one entry point with a website/PIF toggle, instead of three separate buttons/pages
Richer AI extraction: description and past-experience prompts now ask for real detail (one paragraph per named project, not one compressed sentence) instead of a generic summary, since domain/technology auto-tagging is grounded in exactly this text
Weighted matching: AI-scored matches now blend the AI's semantic judgment with a deterministic domain/technology tag-overlap score (65% AI / 35% tag overlap by default), with the breakdown shown in the rationale — the "yes/no eligibility gate, then a weighted score" structure made explicit and transparent instead of one opaque AI number
Review-profile page redesign: grouped into Identity / Description & experience / Domains-technologies-expertise sections, 2-column layout for compact fields, and a live "selected tags" summary bar above each picker (with one-click removal) so you don't have to hunt through ~30 collapsed domain categories to see or edit what's already chosen

v0.2 (shipped)

Deadline-urgency status — Expired / Act now (≤30d) / Prepare now (30-90d) / Future (90d+), computed live against today's date, plus a manual Archive override
Opportunities list: checkbox filters for Status, Matching level (0-20/20-50/50-70/70-100%/not-yet-scored) and Source, a free-text Search box, and Sort (deadline/match%/budget/newest)
RSS/Atom feed support for sources, as an alternative to HTML-listing crawling — more reliable candidate discovery on sites that offer a feed
"Test source" dry run: verifies a source is configured correctly (finds candidates, confirms a sample look like real calls) before committing to a full, possibly multi-minute "Check now"
LAN access — reachable from other devices on the network, not just this machine
Matching-engine fix: eligibility checks now understand synonyms (non-profit/NGO, SME, EU member states, etc.) instead of requiring literal text matches — a profile could previously score 0% on everything if its org type/country was phrased differently than the source text

v0.1 (shipped)

Manual "Check now" ingestion from monitored sources, with a live progress bar
Five seeded sources: Cyprus Funding Programmes Portal, EU Funding & Tenders, Cascade Funding Hub, PRIMA, Interreg Euro-MED
Ollama (local) AI provider, with Claude/OpenAI available in Settings
Human-in-the-loop organization profile creation from a multi-page crawl (homepage + About/Team/Projects-style pages) synthesized into an AI research brief before structured extraction
Hard-filter (eligibility, budget, call type, duration, intensity) + AI Matching% scoring
Manual add-opportunity-by-URL
Background jobs for every crawl/extraction step, so the UI never blocks