fix: implement debounced barge-in check, greeting isolation, and remove aggressive dynamic VAD fast-tracking to eliminate premature filler interruptions and cut-off greetings

fix: inject final critical overrides at absolute bottom of system prompt to enforce behavior rules over database configs

fix: implement robust state barge-in transitions, fix safety reset duration check, and upgrade conversational behavior guidelines

fix: resolve moaning VAD fillers and prevent resuming old context on new calls

fix: 3 critical bugs — choppy audio (batch TTS), hello-hello (silence timer reset), instant VAD filler

feat: complete latency tracking, API retries, echo suppression, and session watchdog

feat: parallel TTS, pre-synthesized greeting, edge cases, dynamic VAD

fix(CRITICAL): correct audio format mismatch - VAD outputs float32 PCM not mulaw, causing all transcriptions to be garbage/empty

fix: VAD singleton + timer-based listening reset + wrapped init in try/except

fix: implement barge-in grace period and increase VAD threshold to prevent echo-interrupts

perf: implement global connection pooling for sarvam ai to fix high latency

fix: resolve backend build dependencies and stabilize Sarvam integration

feat(E5): complete Gujarati perfection — Sarvam STT, fixed TTS, code-mix prompts

fix: personalized greetings and filtered technical tags

feat: Smart tool results and Mock Memory for Demo

fix: resolve NameError List in calendar_service.py

feat: Maya Intelligence & Stability - Smart Availability, Rich Context, and Structured Notes

fix: Full audit round 4 - verbose STT logging, Devanagari Hindi for pronunciation, prewarmed calendar phrases, float32->int16 WAV fix verification

hotfix: Revert English voice to tanya for stability

hotfix: Fix English TTS voice to niharika (bulbul:v3 compatible)

fix: Rewrite STT (float32->int16 WAV conversion + correct model name) + better TTS voices

fix: Full audit round 3 - speechbrain graceful fallback, STT echo bug (remove prompt param), UUID guard in caller_memory

hotfix: Fix STT initial_prompt crash + datetime import + Hindi pronunciation (hun->hoon)

fix: Full enterprise audit - 6 critical bugs fixed (UUID spam, LLM silent bug, tool call accumulation, type safety, default CallerProfile, emotion serialization)

hotfix: Update deprecated speaker voices for Sarvam Bulbul V3

hotfix: Remove deprecated pitch and loudness from Sarvam TTS payload

feat(E4): Caller Memory and Real-Time DB Lookups

feat(E3): Google Calendar Real-Time Booking via Groq Function Calling

feat(E2): Emotional Intelligence Engine + Manual WhatsApp Alerts API + Scaling Roadmap

style: changed gujarati voice to female (vidya) for consistency

feat(E1): neural Silero VAD + Sarvam TTS + language-aware silence

feat: structured note-taking with JSON state extraction and incremental updates

fix: extreme reliability upgrades (llama 8b, long timeouts, and localized static fallbacks)

fix: defined system_prompt in streaming method and switched to stable llama-3.1 model

fix: emergency audio flow restoration (fixed VAD filter and agent state timeout)

feat: Proactive Smart Upgrades (VAD, Hallucination Filter, Abbreviation Handling, Sanitization)

fix: automatic bypass, continuous language switching, hindi default, and removed annoying fillers

feat: Bland.ai/Retell.ai grade — streaming LLM-to-TTS pipeline, echo prevention, language lock fix

feat: add VAD, barge-in interruption, filler sounds, and async pipeline

feat: enhance voice quality, language detection, and reduce latency

fix: update deprecated Groq model IDs to prevent LLM crash and silent calls

fix: graceful handling of missing GROQ_API_KEY to prevent container crash on startup

fix: move server.py to app.py in root to satisfy HF scanner

fix: scrub pycache and env files to clear abusive flag

deploy: flattened structure for Hugging Face