Replace local LLM with HF Inference API + Pollinations fallback - AI chat now works on HF Spaces without model download

Force AI mode always available: on-demand model loading trigger, optimistic availability

Switch LLM from TinyLlama-1.1B to Qwen2.5-0.5B-Instruct (500MB, faster, better answers)

Fix AI mode: fall through to rules when TinyLlama unavailable instead of showing error

Fix math evaluator: catch bare expressions like 2+2, 15*7, 2^10

Add smart chatbot without LLM: math evaluator, 13 general knowledge topics, better fallbacks

Fix AI chatbot loading state: non-blocking model check, loading/failed status tracking, status polling UI, better fallback messages

Replace experience-based interview questions with real coding/debug/arch problems + fix chatbot AI fallback and mode label

Fix chatbot word-boundary matching, add skills/topics, improve AI replies

Replace broken SLM/LLM with TinyLlama for chatbot: rules first for platform Q, TinyLlama fallback for general knowledge, math, coding

Fix chatbot matching: remove word-overlap scoring (caused false matches), add salary/pay keywords to trending topic, improve phrase-only matching

Fix chatbot replies + submit-answer crash + error handling: fix prompt formats for distilgpt2/DialoGPT, move AI state to INTERVIEW_SESSIONS, add try/except on all endpoints, default chatbot to rule mode, deduplicate Socket.IO

Fix security, bugs, consistency across 10 files: XSS, hardcoded IDs, thresholds, dead code, Docker ports

Overhaul chatbot: 30 conversational topics, system-prompted AI, phrase-priority scoring

Add NER blacklist + require 0.7 confidence for unknown terms

Overhaul resume skill extraction: 332 keywords, 153 aliases, section-aware, 25-skill limit

Upgrade interview system: LLM-as-judge scoring, context-aware questions, detailed breakdowns

Fix werkzeug error and cache dir for HF Spaces

Fix python-magic for Linux

Fix flask-socketio version, drop allow_unsafe_werkzeug

Fix colorFrom metadata

Initial deploy