Feature journey
67 levels shipped in 19 days.
Every capability in the order it landed, from the first scaffold to the report that ties the loop together. Nothing here is a promise; all of it shipped.
- days with a release
- 15
- turning points
- 7
- kinds of work
- 7
01Day by day
Every level, on the day it landed.
The busiest day shipped 9 levels, on Aug 19. Tap any dot to read it.
Point at a dot to name it
02Turning points
7 moments the product changed shape.
Most levels are steady work. These are the pivots: a new surface, a revamp, a rule that every later level leans on.
03Every level
The whole log, in the order it shipped.
What each level was for, and what it made possible next.
Aug 8
- L01
Product foundation
The project's skeleton: one workspace holding the web app and the API before either had real functionality.
- Single workspace for every app
- Web app scaffold
- API scaffold
- Shared code standards
- L02
Sign-in foundation
Replaced an earlier sign-in integration with the identity layer every later permission check and data-isolation rule would build on.
- Session handling
- Login & session components
- Configurable sign-in
- Foundation for all later roles and permissions
- L03
Shared UI kit
A shared component library and brand theme, extracted early so every screen built afterward shares one visual language.
- Shared component set
- Brand theme
- Reusable components across apps
- Base for command palette & site UI
- L04
Navigation & theming shell
Replaced the placeholder dashboard page with real sidebar navigation and the first brand palette, giving the app an actual shell to hang pages on.
- Sidebar + breadcrumbs
- Themeable colours
- First brand palette
- Interview loop & course management nav entries
- L07
Course management + interview loop UI
The first real product screens beyond scaffolding: browsing courses, managing clients, and viewing interview details.
- Browse courses page
- Client management page
- Interview details page
- Reusable dialog/breadcrumb components
- L01
Aug 9
- L05
Course builder v1 schema
The first data model for authoring courses, with 14 content item types, built fast and reshaped twice within days.
- Courses, chapters and items
- 14 content item types
- Owner-only access
- Course asset storage
- L08
Chrome extension bootstrapped
The Chrome extension is created as its own app, the seed every capture, permission, and upload feature grows from.
- Extension scaffold
- Separate build pipeline
- Foundation for all capture features
- L09
Social login
Sign-in expands past email and password to Google, Microsoft, GitHub, and Apple, widening the signup funnel ahead of public launch.
- Google/Microsoft/GitHub/Apple OAuth
- Auth callback redirect handling
- Social sign-in on the login page
- L10
Transcription service bootstrapped
A dedicated transcription worker, deliberately separate from the request/response API, converts recordings into timestamped transcript segments.
- Audio extraction
- Speech-to-text transcription
- Standalone background service
- L11
Meeting detection
The extension's first way of knowing a meeting is actually happening, deriving a meeting key and platform from the tab URL and title.
- Meeting-key / platform derivation
- Meet, Teams and Zoom link matching
- Foundation for meeting-end detection
- L12
Course builder rebuild
Hardened against a missing schema and rebuilt as a proper internal authoring tool, replacing the first course-builder page within days of shipping it.
- Course builder workspace page
- Create course dialog
- Trigger & schema hardening
- L05
Aug 10
- L15
Chunked recording + durable storage
Recorder output is written to the browser's local storage in ordered chunks, so a recording survives a crash before it ever has anywhere else to go.
- Chunked recording
- Durable local storage
- Serialized, ordered chunk writes
- L15
Aug 11
- L06
Interview loop schema
The hiring-pipeline data model everything downstream depends on: clients, job templates, and interviews, matched to a meeting by a derived key.
- Clients, job templates and interviews
- Meeting link matching
- Foundation for extension hand-off
- L14
Microphone permission gate
Because the browser asks for microphone access per site, the extension walks the user through granting it on the meeting tab before anything can record.
- Guided microphone permission
- Denied / blocked state handling
- Precondition for capturing candidate voice
- L16
Upload outbox with retry
Every take uploads from its local copy through an outbox, so a dropped connection is retried rather than losing the recording.
- Upload from the local copy
- Retry on a backoff
- Upload on every take, not a setting
- L06
Aug 12
- L13
Client / interview / job template creation
Hiring data stops being read-only: full forms for creating clients, interviews, and job templates.
- Client/interview/job-template forms
- Client logos
- Job template editing
- L17
Server-side audio ingest + transcode
The backend can now receive chunked uploads, reconstruct them byte-exact, validate them, and transcode to a clean audio file.
- Chunked recording ingest
- Byte-exact chunk reconstruction
- Media validation
- Lossless transcode pipeline
- L18
Diarization + transcription worker
Transcripts learn to tell speakers apart: a diarization pass is added alongside speech recognition, plus the first worker/scheduling capability.
- Speaker diarization
- Word-level speaker merge
- Background worker
- Transcript saved with the interview
- L19
Media player + preferences
Recordings become watchable inside the app, alongside the first theme-management pass and a user-preferences provider.
- Media player integration
- Theme management
- Preferences provider
- L20
First analysis pass
The system scores an interview for the first time, an early transcript-analyzer endpoint feeding a first-draft competency display.
- Transcript analysis endpoint
- Early competency scoring
- Analysis display in the UI
- L21
Semantic question search
Interview questions become searchable by meaning rather than exact text, via vector embeddings and a similarity-search function.
- Meaning-based question matching
- Similarity search
- L22
Interview to extension hand-off
A recording finally learns which interview it belongs to: the meeting key is resolved to an interview server-side.
- Meeting-key to interview resolver
- Speaker hand-off to the API
- L23
Mandatory recording + extension UI rebuild
Recording and upload stop being user-toggleable settings, and the popup is rebuilt to match.
- Mandatory-recording policy
- Popup rebuild
- Legacy toggle settings removed
- L13
Aug 13
- L24
Silence detection
The rebuilt popup warns, per take, when the mic or tab audio is actually empty, catching a dead recording before it's mistaken for a real one.
- Per-take silence detection
- Popup silence-advice component
- L25
My courses
Learners get their own dashboard for the first time, distinct from the authoring canvas, connected directly to enrollments.
- My courses page
- Enrollment connection
- First learner-facing surface
- L26
Meeting-end detection engine
Replaced fragile banner-text matching with weighted multi-signal presence scoring and a checked recording state machine.
- Multi-signal presence scoring
- Debounced join/leave detection
- Recording state machine
- Reliable stop-on-meeting-end
- L27
Crash / session recovery
When the browser evicts the extension mid-recording, orphaned sessions are now detected on wake and marked interrupted instead of silently lost.
- Orphaned-session detection
- Interrupted-state recovery
- L28
Competency score locking + caching
Scores stop flickering between page loads: once computed, competency scores are locked and analysis results are cached.
- Competency score locking
- Analysis result caching
- L24
Aug 17
- L29
Recording-gated meeting entry
The meeting itself is now gated behind an active recording, not just a microphone grant, and short failed takes are handled gracefully.
- Recording-gated join flow
- Short-take failure handling
- L29
Aug 19
- L30
RBAC core
Permissions become data instead of hardcoded roles, with scope-aware checks.
- Roles & permissions as data
- Scope-aware permission checks
- Org unit management tree
- Seeded system roles & permissions
- L31
Audit log
Every permission and role change becomes traceable, shipped in the same wave as RBAC core itself.
- Audit log
- Change tracking for role and permission changes
- L32
Server-side permission checks
Every sensitive route now checks the caller's permissions itself, on the server, before it does anything.
- Server-side permission checks
- Roles enforced on every API route
- L33
Roles & permissions settings UI
Admins can now manage who can do what without touching the database directly.
- Role editor
- Permission detail views
- Settings page integration
- L34
Server-checked interview authorization
Before recording starts, the extension now asks the API whether this meeting maps to an authorized interview, and fails closed on any error.
- Authorization lookup check
- Fail-closed on error/timeout
- Short-lived session cache for denials
- L35
Filename-from-meeting export naming
Recordings are finally named after what they actually are, a sanitized, timestamped filename built from the meeting title.
- Sanitized meeting-title filenames
- Local-timestamp suffix
- L39
Deterministic analysis context
A single place now decides exactly what the model is allowed to see, eliminating stale or partial context bugs.
- Single context-assembly point
- Eliminated partial-context bugs
- L40
Transcript metrics + job rubric resolution
Scoring grounds itself in computed fact instead of inference, with a real weighted job-template rubric replacing a keyword fallback.
- Deterministic Q&A metrics
- Weighted skill rubric
- Removed dead keyword fallback
- L41
Analysis as background job
Analysis moves off the page-visit request path onto a queued worker, so opening an interview no longer blocks on a fresh scoring run.
- Queued analysis worker
- Non-blocking UI during analysis
- L30
Aug 20
- L36
Multi-tenancy core
The single-workspace product becomes a real multi-tenant one, with organization identity retrofitted across nearly every business table.
- Tenant / organization / membership model
- Organization identity on all business data
- Invite path for people outside the domain
- L37
Org-scoped data isolation
Isolation stops being modeled and starts being enforced: existing data is backfilled, and access policies are activated everywhere.
- Backfilled organization data
- Org-aware permission functions
- Activated isolation policies
- L38
Org switcher + management UI
Users who belong to more than one organization can now move between them, and admins get a real management screen.
- Sidebar org switcher
- Org settings / management screen
- L42
Analysis history + status lifecycle
Every analysis run is now kept rather than overwritten, with an explicit status lifecycle backing it.
- Analysis history table
- Pending/status state tracking
- L45
Grounded insights + quote verification
Every quote the model cites is now checked against the actual transcript, with an honest low-evidence verdict replacing fabricated scores.
- Evidence verdict system
- Quote verification against transcript
- Structured evidence citations
- L46
Confidence scoring from evidence
Confidence stops being the model's opinion of itself. It's computed from quote volume, directness, and audio quality.
- Evidence-based confidence calculation
- Unanswered question treated as a strong gap signal
- L36
Aug 21
- L43
Email invite ingestion
Interviews can now be created straight from a calendar invite in email, deterministic parsing plus a narrow AI pass for names buried in free text.
- Inbox-based ingestion
- Calendar invite parsing
- Candidate matching
- L44
Secondary mailbox transport
A second mailbox provider joins the first under a shared, transport-agnostic contract.
- Secondary mail transport
- Shared mail-source protocol
- L47
Recommendation reasoning
Hiring recommendations now come with real, generated reasoning rather than a bare verdict.
- Recommendation reasoning generation
- Inferred insight depth
- L48
Interview report card set v1
The first real interview report UI: decision, coverage, strengths, talk time, timeline, and evidence.
- Decision card
- Interview coverage view
- Strengths / watch-outs card
- Talk time, timeline & evidence cards
- L43
Aug 22
- L49
Marketing site launch
The product gets a public face: home narrative, product deep-dive, pricing, and a demo-request flow.
- Home / product / pricing pages
- Demo request form
- Security & legal pages
- L50
Command palette v1
A first launcher gives cross-app navigation a keyboard shortcut, ahead of the deeper rebuild it would get later.
- Keyboard launcher
- Basic command search
- L51
Capability analytics foundation
A read-only observability layer over the AI pipeline is born, isolated so an analytics bug can never take the core product down with it.
- AI quality domain
- Interview intelligence domain
- Non-blocking reconciler worker
- L52
Interview report card set v2
The interview report grows to eleven cards: concerns, gap to learning, unanswered questions, follow-ups, and a closing summary.
- Concerns card
- Gap to recommended learning card
- Asked & not answered card
- Follow-up questions & summary
- L53
Coaching guide generation
A scored gap now comes with what a strong answer would have looked like, not just a score.
- Expected-answer generation
- Coaching guide content
- L54
Tenant onboarding + branding
New workspaces get a guided first-run wizard, including logo and accent branding, ending in a route to the org's first interview.
- Onboarding wizard
- Logo & accent branding
- Workspace preview
- L49
Aug 23
- L55
Evidence-grounded course recommendations
Scored gaps translate directly into recommended courses, grounded in the same evidence chain as the score itself.
- Course recommendation engine
- Evidence-linked suggestions
- L56
Decision card redesign
The hiring decision gets one clear, unified presentation with a single score and deduplicated gaps.
- Unified score presentation
- Hiring-fit label
- Gap deduplication
- L55
Aug 24
- L57
Analytics rebuilt around capability
The legacy assessment-plumbing framing is replaced with a role-aware analytics home that answers a real question, not just what ran.
- Role-aware analytics home
- Capability-first framing
- L58
Learning assignment ledger
A recommendation becomes an auditable ledger: assignments carry the source interview, the gap, a due date, and a full waiver trail.
- Source interview & gap linkage
- Due-date tracking
- Waiver trail
- L59
Grading infrastructure
The learning loop actually closes: quizzes are graded server-side, learners can hand in work, and reviewers get a real grading queue.
- Server-graded quizzes
- Learner submissions
- Grading queue for reviewers
- L60
Learning / studio domain split
Authoring and learning finally get separate homes, with every old link still redirecting correctly.
- Learner surface
- Authoring surface
- Legacy link redirects
- L61
Certificates + accessible player
Finishing a course now produces something to keep, and the course player works without a mouse.
- Certificate generation
- Keyboard-navigable player
- L62
People directory + profile
Every person gets one page spanning both their interviews and their training.
- People directory
- Person profile page
- Competency trend
- L63
Global search, work queue & notifications
Real operational visibility replaces three placeholders: a working search box, a real work queue, and a full notification history page.
- Working global search
- Real work-queue dashboard
- Notification history page
- L64
Bulk list operations
Interviews and clients get search, filter, sort, and bulk re-analyze, so stale scores can be refreshed in one action.
- Search / filter / sort on lists
- Bulk re-analyze action
- L57
Aug 25
- L65
Command palette rebuild
The palette stops being a search box and becomes a real action layer, with recents, scopes, fuzzy matching, and page-aware actions.
- Recents & scopes
- Fuzzy match with highlighting
- Page-aware drill-in actions
- L66
Animated AI-insights loading state
Waiting for an analysis to finish becomes legible instead of a blank spinner.
- Sparkle-cluster loading animation
- Auto-advancing step indicator
- L65
Aug 26
- L67
Interview report unification
Three generations of report cards become one coherent report, with real tabs and cross-section links.
- Unified report tabs
- Cross-section links
- Consolidated, de-duplicated cards
- L67
04Your turn
Bring one lost round.
See it scored.
Twenty minutes with your delivery head or staffing lead. Leave with a scorecard, the quotes behind it, and the course that closes the gap.