Release notes
What's new in EzRecruit
The Recruitment Agency Operating System — Streamlined operations, better submissions.
-
Global search is now disabled by default for new tenants
When a new tenant is set up, the global search capability is switched off unless explicitly enabled. This gives administrators deliberate control over which tenants have access to the feature rather than having it active automatically.
-
Duplicate CV batch upload registration removed from API
A copy-paste leftover caused the bulk CV upload endpoint to be registered twice in the API, producing a duplicate entry in the API's auto-generated documentation and potentially confusing any tools or integrations that consume it. The redundant registration has been removed; the endpoint itself is unchanged and continues to work as before.
-
Industry classification results are now shared across servers and survive redeployments
When EzRecruit classifies a candidate's role or company into an industry, that result is now stored in a shared database cache rather than discarded on each restart or across parallel servers. This reduces repeated AI calls for the same lookups, lowering cost and speeding up CV processing at scale.
-
Automated test checks now run before every release build
The full test suite (219 tests) was previously never executed automatically, meaning code issues could reach a release undetected. Every build now runs all tests and will block a release if any fail.
-
Job description upload and paste now reach the AI parsing service correctly
Both the paste-text and file-upload paths for job description parsing were being rejected before any processing occurred, meaning no JD submitted through the normal UI was ever parsed. The routing layer now correctly handles both input types and directs each to the appropriate downstream endpoint.
-
Special characters in AI-generated chat answers no longer appear garbled
Streamed responses containing em-dashes, accented letters, or emoji were being corrupted before reaching the user and being stored in chat history. The encoding mismatch has been corrected, so all characters now display and save correctly.
-
Candidate ranking no longer silently drops scores for some candidates
When ranking a batch of candidates, up to 4 out of 15 could come back with no score, strengths, or gaps shown — silently discarded with no error. The root cause was the AI model occasionally shortening a candidate's identifier, causing the score lookup to miss. Candidates are now labelled with simple positional references during ranking so the model never needs to reproduce a long identifier, and any remaining mismatches are logged as warnings rather than dropped.
-
Automated code-review safeguard now consistent across all repositories
The override permission used to bypass a blocked automated review has been aligned across two additional repositories, matching the safeguard already standard elsewhere in the organisation.
-
UAT deployments unblocked after a day-long pipeline outage
An automated dependency update introduced a version of a machine-learning library that is incompatible with the Python version this service runs on, causing every UAT build to fail for roughly 24 hours and preventing any changes from reaching the UAT environment. The dependency has been pinned back to the last known-good version, restoring the pipeline. Upgrading to a newer version will be done deliberately as a separate piece of work.
-
Service credentials strengthened across build and deployment pipelines
Eight service credentials, including database, search, and AI API keys, have been moved out of build artefacts into runtime-injected secret storage, extending a protection standard already applied elsewhere in the platform. Four additional credentials were moved to properly secured storage in the deployment pipeline, verified to match previous values before going live.
-
Candidate ranking no longer silently drops candidates when the AI returns a malformed response
In live UAT traffic, the AI model occasionally returned a candidate identifier with extra text run into it, which the ranking logic did not recognise, causing that candidate to be left unscored. The fix detects this specific malformed pattern and recovers the correct identifier. Previously the failure was silent; it only came to light because earlier logging improvements made unranked candidates visible.
-
Global Search now streams live progress to the UI as results are found
The Global Search screen was designed to show step-by-step progress as a search runs, but the backend was returning a single blocked response only after all work was complete, causing the server to reject the streaming request entirely with an error. The endpoint now streams results progressively to the frontend as they arrive, matching the intended user experience.
-
Automated evaluation framework can now measure whether a configuration change affects search results
A baseline of expected search results has been generated from 58 real recorded conversations, enabling the team to run future configuration changes against it and detect any drift in candidate results. This unblocks a pending decision about a caching feature that could not be evaluated without a reference baseline.
-
Critical framework vulnerability patched; credential handling hardened
The image-handling framework used across the application has been updated to the latest patched version, addressing a publicly disclosed vulnerability. Separately, service credentials embedded in container images have been moved to runtime-only injection, and a previously incomplete credential rotation has now been completed.
-
Candidate search results now paginate, dropped connections can be resumed, and error messages are plain English
Recruiters can now page through all matching candidates beyond the first 20 results, with a clear count of total matches shown. If a network drop interrupts an in-progress search, a Reconnect button lets the session resume from where it left off rather than losing the results entirely. Error messages throughout the search experience have been rewritten in plain language instead of displaying raw technical codes.
-
Goal settings page no longer crashes on server-side load for users with no received goals
Users navigating to the received-goals settings page when they had no goals assigned were triggering a recurring server-side crash that had been occurring every few minutes since June. The page now loads correctly in all cases.
-
AI usage and cost-tracking data is now reliably recorded for all tenants
Five database tables used to track AI call attribution and per-tenant AI balance and quota were missing from the standard tenant setup process, meaning UAT had none of them and production was missing the balance tables entirely. Calls continued to work but attribution rows and balance lookups were silently failing on every call. These tables are now part of the standard tenant provisioning flow, closing the gap between environments and ensuring cost and usage data is captured correctly going forward.
-
Email delivery can now be switched to a lower-cost provider without downtime
The application has been prepared to send emails via a second provider at a significantly lower cost (approximately $0.10 per 1,000 emails versus a $35 flat monthly fee). The existing email provider remains active; no sending behaviour changes until the switch is explicitly enabled.
-
Users no longer lose work after a production deployment
After every deployment, users who had the application open in a tab would receive an error on any action — submitting a form, saving a record — until they manually refreshed the page. This was caused by a security key being regenerated on every build, making the newly deployed server unable to recognise requests from pages loaded before the deployment. The key is now fixed and stable across all future builds, so this class of disruption will not recur.
-
Missing AI usage and quota tables restored for affected tenants
One tenant was missing four database tables required for AI balance and quota tracking, and all tenants were missing a user identifier column on AI usage records. Both gaps are now corrected in production. The changes are safe to apply repeatedly and have no effect on tenants whose schemas were already up to date.
-
Email delivery can now be migrated to the new provider one notification type at a time
Previously, switching email infrastructure required moving all notification types at once. A new configuration option now allows individual notification types — welcome emails, OTP codes, interview invites, and so on — to be moved to the new provider independently, while everything else continues on the current provider. This enables a controlled, low-risk rollout with the ability to pause or reverse at any step.
-
Automated code reviews now flag issues that were already raised in a previous round
The automated pull-request review tool can now detect when a finding it raises was also raised in an earlier review of the same change. Repeated findings are labelled as such and counted separately, making it easier for reviewers to see which problems have persisted across revisions rather than treating every round as a fresh list.
-
Post-interview messages and reminders now send reliably after a 14-day outage
A July code change accidentally removed the database session that these scheduled jobs depend on, causing every single run to fail silently — 219 consecutive failures across all production tenants, with zero post-interview messages delivered in two weeks. The jobs now open a proper transaction per interview, load all required data in one query, and skip a single failing record instead of halting the entire tenant's queue. A data clean-up also marks the backlog of past interviews that can no longer be reminded as resolved, so they stop appearing as outstanding work.
-
External job applicants now receive a correctly branded verification email
Candidates applying for jobs through the public portal were receiving a one-time password email intended for admin users, complete with admin portal branding and a greeting placeholder that was never filled in. The verification email for external candidates now uses its own dedicated template with the correct audience and content.
-
Goal notification emails now send successfully
Goal creation and edit notification emails were silently broken when the email provider requires an explicit subject line. This was caught during pre-launch testing of the new email provider and fixed before it could affect live traffic.
-
Resume uploads no longer fail when credentials are rotated or removed from the build
A recent deployment shipped without the expected credentials baked into the image, causing all resume uploads to fail for approximately 30 minutes. The service now falls back gracefully to the server's attached identity when explicit credentials are absent, so a missing build secret no longer takes the upload feature offline.
-
Server actions remain functional across redeployments in the admin portal
Every deployment of the admin portal was generating a new random encryption key, which invalidated any in-flight server actions from the previous version. The key is now stable across deploys, preventing unexpected failures for users active at the moment of a release.
-
Candidate interview-slot page: bookings no longer fail, declines now work, and error details are hidden
Three in ten candidate self-bookings were failing silently due to a race condition when generating offer IDs, and every attempt to decline an interview returned an authentication error. Both are now fixed. Additionally, raw database error messages that were previously shown verbatim to candidates on booking failure have been replaced with plain, readable text. The 'Not Interested' and 'Request other times' buttons are now also visible even when no slots remain.
-
Interview confirmation message now sends only after a booking succeeds and shows the correct time
Candidates were receiving a WhatsApp confirmation for bookings that subsequently failed, and a failed message send could roll back an otherwise successful booking. The confirmation now only fires once the booking is fully committed. Separately, the interview time shown in the message was displaying 5.5 hours off due to a timezone handling error; it now shows the correct local time.
-
Email delivery can now be routed through a second provider with automatic fallback
The platform can now send notifications through either the existing email provider or a new alternative, controlled by configuration with no code change required. If the new provider fails on a given send, the system automatically retries through the original provider and raises an alert, so no notification is silently lost. Three additional notification types have also been prepared for use with the new provider when needed.
-
Candidate search service restored after database connectivity failures on UAT
All database-backed search routes were returning errors on UAT because the database connection was being made without the encryption the server requires. Encrypted connections are now enforced, restoring search functionality. Internal database calls within the search pipeline were also aligned with the service's asynchronous design to ensure consistent, predictable behaviour.
-
Rolling back a deployment no longer fails when the target version is already live
Re-running a rollback to the currently deployed image previously failed partway through, causing the forced redeploy step to be skipped. The process now recognises that the image tag is already correct and proceeds to trigger the redeploy as intended.
-
Latent defects closed before clients hit them
Two crashes in the reporting endpoints, a conversion ratio correct only by accident, a permanent guard against the startup crash recurring, three fields the interface expected but never received, and code linting re-enabled after having been switched off.
-
Suspended clients now receive database updates
Ten of twenty-three client databases were stranded up to 140 versions behind. Reactivating any one of them would have run the entire backlog against live traffic.
-
Emergency rollback now reliably finds all production services
A limit in how cloud services are listed meant that four of the twenty-three services — including the main production backend — were invisible to the rollback process. Had a rollback been triggered during an incident, it would have failed silently at exactly the wrong moment. All services are now found correctly regardless of how many exist.
-
API documentation on production restricted to internal use
The full internal API reference, covering 582 routes with parameters and data schemas, is no longer served on production. Development and staging environments retain access for internal teams, tightening the boundary between public and internal surfaces.
-
Security vulnerabilities patched in web framework and linting configuration
The web framework used by the AI routing service was updated to address a known session-handling security advisory. A YAML parsing vulnerability in the admin portal's linting configuration was also resolved. Both updates eliminate publicly disclosed security risks.
-
Credentials now fully protected in build pipelines
Database passwords and third-party API keys used in build pipelines have been moved to encrypted secret storage, matching the protection standard used elsewhere in the platform. Automated deployment to the UAT environment is also restored.
-
AI cache usage is now accurately recorded, enabling cost and effectiveness reporting
AI calls served from cache were being billed correctly but logged as if no caching occurred, making it impossible to measure whether caching was working or saving money. Both paths that write billing records now capture the actual cache token counts, so reports and future decisions about caching strategy will reflect reality.
-
Independent model selection for JD parsing
JD parsing was locked to one setting at ₹0.66 per call with no way to tune it, plus a quality guardrail added on ranking.
-
Session expiry, tenant links and permissions
Users with an expired session saw a raw system code instead of the login page; admin portal tenant links rendered as broken URLs; and Goal Settings demanded create-level permission on a read-only screen.
-
Candidate record growing without limit
A candidate's technology field grew each time they were tagged to a requirement, with no cap, eventually overflowing and crashing evaluation submit.
-
Four billing defects in AI call settlement
Each silently lost or double-counted revenue rather than producing an error, which is why none had ever been reported. Also fixed: US-region résumé file paths were being corrupted.
-
Human approval required before production
Automation had produced eleven unreviewed automatic production releases in a single day. An explicit approval is now required, and direct pushes to production are blocked — everything must pass through UAT first.
-
Failed database updates are now recorded
Previously a failure left no trace whatsoever and retried silently forever. The single biggest visibility gain of the period — and what surfaced the two items below.
-
AI spend controls
Monthly budget per client with per-person and per-feature allowances, a Super Admin panel to set them live without a release, a client-facing usage dashboard, and every AI call attributed to the employee who triggered it.
-
Email module revamp
Custom tracker and full redesign — the largest single product change of the period.
-
Interview roster filters that actually filter
Seven filters were shown in the interface but ignored by the system. Those now work, along with column sorting and Recruiter Name in both the grid and the Excel export.
-
"Failed to fetch organisation config" at login
Resolved, and the underlying database update process re-enabled after having been switched off since early August. Three client workspaces that could not receive any further updates were unblocked at the same time.
-
Application crash on startup blocking every release
A duplicate component name crashed the application at startup, so every release was automatically rolling back and no new code could reach production.
-
Error on every signed-in request for India-region clients
A database query was reading a field it never requested. Masked until now by the startup crash above.
-
Requirement editing and long skill lists
Editing older requirements crashed outright, and long skill lists were rejected on save.
-
"Failed to load filter options" across nine modules
Any client without location data saw an error rather than an empty list, affecting nine separate modules.
-
AI cost reporting and missing billing tables
Two production tables were missing entirely, JD parsing and search-query generation were reported as a single cost line hiding each one's real cost, and queries were hardened against schema drift.
-
CV parser reading its own labels as company names
Eighteen internal section labels — "Fintech", for example — were being extracted as real company names on every CV.
-
AI Search quota message, Global Search hidden, health check false alarm
Users hitting an AI Search limit saw nothing because the message went to a component that no longer existed; Global Search was hidden from the menu while still in build; and a health check was reporting a false alarm on a healthy service.
-
Production credentials moved out of plain text
Credentials across six services moved into masked storage. Previously every collaborator could read them.
-
AI Search monitoring
Deep health checks, encrypted-connection verification, pinned dependencies and two production alarms. The service previously had none.
-
Single active session per user
Signing in now evicts the previous device, so one account cannot be shared across people simultaneously.
-
Multiple collaborators per supervisor
A supervisor can now have several active collaborators attached at once, selected together rather than one at a time.
-
Bulk résumé import
Batch upload and parsing of résumés, rather than one file at a time.
-
Interview roster moved to server-side filtering
Filtering, sorting and paging now happen on the server, so the roster stays fast as record counts grow.
-
Document conversion no longer jams the parser
A single malformed Word document could permanently hang the conversion service, taking down CV parsing for everyone on that instance. Now isolated per document, with the monitoring alarm repaired and extended.
-
Candidate form usability
Image zoom, a logo link back to the dashboard, a status picker, validation feedback on tags, and current-designation autocomplete.
-
Analytics and ROI reporting
The analytics module and its ROI enhancement, including the menu and permission wiring to reach it.
-
Global Search foundation
The menu entry, permissions and proxy layer for Global Search. The feature remains in build and is not yet live for clients.
-
Critical security findings closed in the parsing service
Path traversal, credentials being written to logs, an injection route into the parsing function, and personal data appearing in log output — all fixed together.
-
Autocomplete, document handling and module routing
Supporting-document input handling, designation autocomplete with exact matches ranked first, and module URL segregation behind a proxy layer.
Nothing matches those filters.
