{ Product Progress & Changelog }

ROADMAP & RELEASE NOTES

Phases in plain words, current development status, and detailed release notes. No dates promised — every capability is built and verified with honest criteria.

> PHILOSOPHY:
> We ship working software with verifiable guarantees, zero tracking, and no lock-in.
> Every phase has an explicit exit gate, automated accessibility tests, and zero dead code.

1. Roadmap by Phase

Abhyas is developed in structured phases, each building upon verified contracts and rigorous client-side privacy architecture.

Phase 0

Foundations & Safe Architecture

✓ Completed

Core contracts, offline-first storage engine, design tokens for Ink and Paper themes, self-hosted fonts, and zero-telemetry boundary enforcement.

  • Strict type safety and boundary checking
  • Ink and Paper high-contrast WCAG 2.2 AA themes
  • Dexie IndexedDB storage engine with validation and quarantine
  • Network egress allowlist and automated license audit
Phase 1

Core Interview Loop & Public Launch

✓ Completed

Full voice and text interview practice with push-to-talk, real-time transcription, schema-constrained evaluation, grounded rewrites, and public documentation.

  • Multi-provider direct client adapters (Gemini & OpenAI-compatible)
  • Client-side document parsing in Web Worker (PDF, DOCX, TXT) with 0 network calls
  • Live interview focus screen with timer, subtitles, hint, and follow-up loops
  • Grounded rewrite generator with anti-hallucination verification
  • Prerendered static public pages with zero external scripts
Phase 2

Skills Dashboard & Habit Engine

✓ Completed

Honest skill statistics, level-calibrated difficulty, practice drills targeting calibrated weak points, daily streaks, and PWA offline installation.

  • 15-skill taxonomy across technical, behavioral, and delivery categories
  • Recency-weighted scoring engine (H=8 half-life) with confidence tracking
  • Interactive skills dashboard with score bands and chronological sparklines
  • Quick-fire and Deep-dive targeted skill drills with 'Fix my weakest' 1-click CTA
  • Offline Service Worker shell pre-caching and installable Web App Manifest
Phase 3

In-Browser & Local Models

✓ Completed

100% offline, zero-network-egress practice powered by WebGPU in-browser models (WebLLM) and local daemon servers (Ollama, LM Studio).

  • Device capability detection (RAM, WebGPU, token decoding micro-benchmark)
  • Strict device stability rules: PC desktop only, text-only enforcement
  • Curated local models catalog with automated post-load sanity testing
  • Local server setup support with classified CORS diagnostic helpers
  • Prominent honesty badge tagging for basic model feedback
Phase 4

Curated Company Context Packs

✓ Completed

Tailor practice sessions with evidenced company context packs covering values, culture, technical stack, and verified interview formats.

  • Evidenced pack contract requiring verified citations, dates, and official sources
  • 12 hand-reviewed seed packs for popular tech employers
  • Pack-aware prompt layering with values weighting and anti-hallucination guards
  • Multi-layer prompt injection defense and delimiter neutralization
  • Company directory and detail browser with offline IndexedDB caching
Phase 5

Demand-Gated Research Service

✓ Completed

Bounded research pipeline to ingest and curate new company packs safely when real user demand passes community thresholds.

  • FastAPI service with strict rate limiting and anonymous client hashing
  • SSRF-safe fetcher blocking private, loopback, and cloud metadata addresses
  • Hard budget caps ($5/day, $100/month) and automated operator kill switch
  • Static mirror export fallback for 100% offline resilience
Phase 6

Preparation Schedule & Spaced Review

✓ Completed

Long-term interview preparation planning with universal .ics calendar integration, ramping difficulty plans, and spaced repetition review.

  • Cadence and target date schedule generation with progressive difficulty ramping
  • RFC 5545 universal .ics export compatible with Google, Apple, and Outlook
  • Client-side Google Calendar integration with in-memory tokens only
  • Spaced repetition review queue with expanding intervals (1d to 60d) and 10/day cap
  • Automatic enqueuing of weak interview turns into review queue
Phase 7

Advanced Providers & On-Device Speech

✓ Completed

Additional BYOK providers, custom local model imports, and local voice activity detection (VAD) for natural hands-free conversation.

  • Anthropic Claude (Messages API) and Hugging Face serverless provider adapters
  • Custom local model support with safe file format checks (.safetensors, .gguf, .onnx)
  • Web Audio API Voice Activity Detection (VAD) for automatic silence detection
  • Expanded 60-question bank with level-calibrated follow-ups and rubric hints

2. Changelog & Release History

All architectural and user-facing changes recorded in progression order.

v0.9.3— Phase 7 Complete: Advanced Multi-Provider & Hands-Free VAD

Added Anthropic Claude and Hugging Face inference support, custom local model validation, on-device voice activity detection, and comprehensive exit gate verification.

  • BYOK direct browser adapter for Anthropic Claude (Messages API) and Hugging Face serverless API
  • Custom local model loading with safety format verification blocking unsafe pickle formats
  • Real-time Web Audio API voice activity detection (VAD) with configurable silence countdown
  • 60-question calibrated bank across all 5 experience level bands

v0.8.2— Phase 6 Complete: Schedule Generation & Spaced Repetition

Introduced prep plans with RFC 5545 .ics calendar export, optional Google Calendar sync, and a daily spaced repetition review queue.

  • Universal .ics calendar generator with 75-octet line folding and strict CRLF formatting
  • Spaced repetition review queue with expanding intervals (1d, 3d, 7d, 14d, 30d, 60d) and 10/day cap
  • Automatic weak turn capture from session scorecard into review queue
  • In-memory Google Calendar event creation with zero persistent credentials

v0.7.6— Phase 5 Complete: Demand-Gated Company Research Service

Launched bounded backend research service with SSRF protection, strict budget caps, and static pack mirror export.

  • FastAPI research service in apps/api with SQLite/PostgreSQL async support
  • SSRF-safe web fetcher blocking cloud metadata (169.254.169.254) and private networks
  • Enforced $5/day and $100/month budget caps with emergency kill-switch
  • Static pack mirror fallback guaranteeing client offline independence

v0.6.5— Phase 4 Complete: Evidenced Company Context Packs

Added company-tailored interview practice with 12 seed company packs, prompt injection safety defenses, and company directory.

  • Evidenced company pack contract requiring verified citations and dates
  • 12 curated seed packs for major employers (Google, Amazon, Meta, Apple, etc.)
  • Values-weighted question selection and company tech stack tailoring
  • Prompt injection defense with untrusted input sanitization and delimiter neutralization

v0.5.6— Phase 3 Complete: Offline & Local Models Engine

Introduced private offline practice with in-browser WebGPU models and local server adapters (Ollama, LM Studio).

  • WebGPU hardware capability benchmark and token decoding throughput speed test
  • In-browser model runner in Web Worker with WebLLM integration
  • Ollama adapter with NDJSON streaming and automated localhost CORS diagnosis
  • PC-only and text-only device safety enforcement with 'Basic Feedback' honesty badges

v0.4.6— Phase 2 Complete: Skills Dashboard, Habit & PWA Shell

Delivered honest skill analytics, category-grouped practice drills, daily streaks with guilt-free rest days, and offline PWA support.

  • 15-skill taxonomy engine with recency-weighted statistics and confidence scores
  • Interactive skills dashboard with horizontal bars and score history sparklines
  • Quick-fire (5 Qs) and Deep-dive (10 Qs) drills with 'Fix my weakest' 1-click CTA
  • Deterministic local habit engine with protective monthly freezes
  • Installable PWA manifest and offline service worker shell

v0.3.8— Phase 1 Complete: Voice & Text Practice Core Loop

Launched complete client-side interview practice flow with AI adapters, speech recognition, rubric evaluation, and public site.

  • Gemini and OpenAI browser adapters with PBKDF2/AES-GCM client encryption
  • Client-side document parsing in Web Worker for PDF, DOCX, and TXT
  • Push-to-talk voice interview with live subtitle lanes and transcript review
  • Rubric-based evaluation with anti-hallucination grounded rewrites
  • Prerendered static public pages with zero third-party network egress