Initiatives we plan to take up in subsequent phases, grouped by theme. Each card has a one-line description, priority and rough effort estimate, and (where one exists) a link to its detailed planning doc. Items already shipped live in CHANGELOG.md — click the version chip on the home page to view in-app.
Features
New product capabilities for recruiters, hiring managers, and candidatesThree-tier support (L1 → L2 → L3 Jira)
Extends the L1 chatbot triage already shipped (v1.2.0). L2 adds a WebSocket live handoff to an IT engineer with the full L1 transcript and tool results attached; L3 escalates to a Jira incident in one click. Closes the loop on "user reports a bug → engineering knows."
Aptitude / logical-reasoning test
Objective cognitive assessment slotted between resume scoring and the shortlist email. 15 questions, 20 min, four categories (verbal, numerical, logical, situational). Adds a third measurable signal alongside resume match and interview performance, and stops Maria from spending a slot on candidates who'd flunk a basic gate.
Async / low-bandwidth mode
Text-only or audio-upload fallback for candidates on 2G/3G connections. Today the connectivity monitor catches poor links but only after the live call has failed. Async would record questions for the candidate to answer in their own time, with async scoring on submission.
Question-bank editor
Recruiter UI to edit interview question banks directly. Today
data/questions/*.json is hand-edited and ships with
the deployment — every change requires an engineer.
Candidate explanation page
DPA Section 16(c) gives candidates the right to know how decisions
about them were made. A /web/my-interview/<token>
page would show them their transcript, dimension scores, and the
high-level reason for the outcome.
Recruiter notes on scorecard
Let the recruiter add their own observations on top of Maria's output. Today the scorecard is read-only; recruiters keep notes in separate spreadsheets, breaking the single-pane workflow.
Bulk re-score
When we improve the scoring rubric, recompute scores for already- closed interviews without re-running the calls. Lets us A/B rubric versions on historical data before promoting changes.
A/B testing different prompts
Iterate Maria's tone and question selection based on completion rate and downstream recommendation quality. Requires a lightweight prompt-versioning layer and per-interview prompt attribution.
Security & Compliance
Auth, data protection, regulatory readinessSSO / Azure AD on admin panel
Closes the #1 high-priority gap — the admin panel is currently publicly reachable. App-level OIDC with HCPH's Entra ID, HttpOnly session cookies, three-group authorisation (Recruiters / HMs / Admins). Feature-flagged rollout.
Privacy notice page
/web/privacy.html covering what we collect, why, who
sees it, retention period, candidate rights (access, deletion,
correction, complaint to NPC). Required for DPA compliance.
Data retention enforcement
Cron job that deletes recordings past the cutoff and anonymises transcripts. Suggested defaults: 12 months for recordings, 24 months for transcripts + scores, 7 years for hire/no-hire decisions (matches Labor Code dispute window). Awaiting policy decision from HCPH legal / DPO.
Opt-out to human screening
Candidate can decline the AI interview and request a human via a "Request human screening" button on the landing page. Today they have to reply to the invite email — works but no formal flow.
Quarterly bias / calibration audit
Sample 5% of completed interviews, have a human recruiter independently score them, compare distributions across age band, gender, region, language. If disparate impact emerges, retune. Process change + a small reporting tool.
NPC registration / DPO sign-off
Confirm HCPH has registered the data processing system with NPC under their existing registration, and that the DPO has reviewed Panayam's processing activities. Org-side action; we provide a checklist.
EU GDPR readiness
Only relevant if international hiring goes live. Article 22 right not to be subject to solely automated decisions — our "AI assists, human decides" framing covers this, but we need to document it formally.
Efficiency & Cost
Finance / OPEX visibility and optimisationCost-per-hire dashboard
Once the hired flag exists on candidates, the
utilization page gains a 5th block: ₱ spent ÷ candidates hired,
per role and overall. Gives finance a defensible ROI number.
Cost block on utilization page
v2 of the utilization dashboard: total ₱ this month, breakdown by provider (LLM / STT / TTS / LiveKit / infra), forecast for next 30 days based on scheduled load. Deferred from v1.
Budget alerts
Cloud Run, Anthropic, ElevenLabs all support programmatic budget alerts. Wire them to a recruiter Slack channel so spend surprises don't happen at month-end close.
Cost per pipeline stage
Today we know cost per completed interview, but not per shortlist, per question-bank generation, per analytics regen. Stage-level cost attribution unlocks better resource allocation.
Vendor swap proof points
Document the Google TTS → ElevenLabs swap we already executed mid-flight as a one-pager for finance — assurance that no single vendor can hold us hostage.
System Integrations
Connections to HCPH's existing HR + collaboration stackDarwinbox integration
Bidirectional sync with HCPH's current HR portal — candidate push from Darwinbox to Panayam, scorecard pull back to Darwinbox. Picks up after POC sign-off.
Greenhouse / Workday / Lever
Same pattern as Darwinbox but lower priority unless HCPH expands to a different ATS. Architecture is the same; only the adapter changes.
Cross-timezone scheduling
The Calendly-style picker shows only Manila time today. International candidates can't easily see their local slot — detect TZ from browser, render both columns.
Slack / Teams alerts
Push notifications to a recruiter channel on key events: candidate completed an interview, scorecard ready, integrity-breach detected. Surfaces work without polling the admin panel.
Scale & Reliability
Throughput, regional coverage, ops hardeningRegional language support
Cebuano, Bisaya, Ilocano, Hiligaynon. Google STT chirp_2 doesn't cover these natively; needs STT model evaluation + prompt translation + rubric validation per dialect. Single biggest demand driver for PAN-Philippines mass hiring.
Concurrency cap lift
Current production cap is 5 simultaneous interviews. LiveKit + Cloud Run can handle more with config changes. Load test at 20, then 50, then 100 concurrent before raising in production.
Multi-region failover
Cloud Run today is single-region (asia-southeast1). Add a passive replica in asia-east1 (Taiwan) or us-west1, with health-check- based DNS failover. Only worth doing once volume justifies it (>200 interviews/day).
Disaster recovery runbook
Document what to do if Anthropic, ElevenLabs, Deepgram, LiveKit, or Cloud SQL has an outage. Today this lives in heads, not on paper.
Public status page
status.panayam.example.com so candidates and
recruiters can see whether the platform is up without having to
test it. Statuspage.io or similar.
Observability via real APM
Push structured logs to Datadog or Grafana Cloud rather than reading the Cloud Logging UI. Set alerts on critical KPIs (interview completion rate <90%, deep-health failures).
Insights & Analytics
Trend reporting, cultural calibration, decision supportTrend lines on analytics
Today shows a snapshot per role. Add week-over-week deltas so the head of recruitment can see if a role's pipeline is improving or degrading.
Cultural calibration of rubric
Maria's rubric is calibrated on Western "self-promotional" norms. Filipino candidates can score artificially low for deference / modesty. Run a human-vs-AI calibration audit on the first 100 mass-position interviews, adjust dimension weights.
Custom rubric library per seniority
Currently every role uses one of two rubric IDs. Add a Director / Manager / IC tier that adjusts question depth, weighting, and Maria's calibration (e.g. expect more strategic-thinking signal at Director level).
UX & Accessibility
Inclusive candidate experienceMobile-first candidate experience
Verify camera/mic permissions, screen-orientation handling, control-bar layout on <400px screens. Mass-position hiring in the Philippines is heavily mobile-first.
Accessibility
Hearing-impaired (live captions already partly exist, formalise + offer text-only mode), blind (screen-reader friendly), speech disabilities (longer VAD timeouts, less aggressive end-of-turn detection).
Source: NEXT_PHASE.md · SSO_INTEGRATION_PLAN.md · APTITUDE_TEST_PLAN.md · DARWINBOX_INTEGRATION.md