import os from dotenv import load_dotenv load_dotenv() NVIDIA_BASE_URL = "https://integrate.api.nvidia.com/v1" # ── Candidate contact ──────────────────────────────────────────────────────── # Resume-header address line. ATS checkers (Jobalytics / Resume Worded) flag a # missing address; a "City, State, Country" line satisfies that check. Used as # the fallback when the uploaded resume has no parseable location. CONTACT_LOCATION = os.getenv( "CONTACT_LOCATION", "Hyderabad, Telangana, India · Open to relocate" ) # ── Maximum ATS Mode (User-Confirmed Skill Expansion) ──────────────────────── # The uploaded resume is a BASE PROFILE, not the candidate's full experience. # When enabled, normal Product/Product-Manager/AI/SaaS/B2B/agile vocabulary is # treated as user-confirmed / interview-supportable and aggressively included # across Summary, Skills, Key Achievements and Experience to maximise external # ATS keyword coverage. This NEVER loosens the hard anti-fake boundaries: # degrees, certifications/licenses, employers, titles, years/seniority and # regulated credentials stay blocked; specialized hands-on eng/security tools # still require explicit confirmation. MAXIMUM_ATS = { "enabled_default": False, # extension toggle controls it per-request "target_external_score": 95, # aim for 95–100 where honestly achievable "min_external_score": 90, # below this (for in-family roles) = failure state "max_repair_iterations": 4, # external-feedback repair iterations } # Normal PM / Product / AI / SaaS / B2B / agile terminology that, in Maximum ATS # Mode, is treated as user-confirmed (LOW risk, always includable) for the # candidate's target role family. These are craft/role terms — NOT credentials, # employers, titles, or specialized hands-on engineering tools. MAXIMUM_ATS_SAFE_TERMS = { # AI / ML product craft (not hands-on model engineering) "ai", "ml", "machine learning", "generative ai", "gen ai", "genai", "llm", "llms", "large language models", "prompt design", "prompt engineering", "cloud ai", "cloud ai service", "cloud ai services", "ai testing", "ai monitoring", "model evaluation", "ai evaluation", "ai strategy", "ai product", "ai product management", "conversational ai", "nlp", "responsible ai", "ai governance", "ai roadmap", # Analytics / experimentation "qa", "quality assurance", "experimentation", "a/b testing", "ab testing", "data-driven", "data driven", "analytics", "product analytics", "metrics", "kpis", "kpi", "okrs", "okr", "sql", "dashboards", "data analysis", "hypothesis testing", "funnel analysis", # Core product management "roadmap", "product roadmap", "product strategy", "product ownership", "product owner", "product discovery", "product vision", "product lifecycle", "go-to-market", "gtm", "prioritization", "feature prioritization", "requirements", "prd", "prds", "user stories", "acceptance criteria", "stakeholder management", "stakeholder communication", "cross-functional collaboration", "cross-functional", "backlog", "backlog grooming", "user research", "customer discovery", "market research", "competitive analysis", "wireframing", "user acceptance testing", # Agile / delivery "agile", "scrum", "kanban", "lean", "sprint planning", "sprint", "agile methodologies", "scrum master", "story mapping", # Business / domain "saas", "b2b", "b2c", "enterprise", "enterprise platform", "startup environment", "startup", "customer value", "adoption", "retention", "growth", "engagement", "conversion", "monetization", "revenue", "customer success", "onboarding", # Cross-functional PM responsibilities & leadership craft (interview-supportable) "influence", "influencing", "program management", "product management", "product marketing", "business development", "global teams", "diverse partners", "customer results", "customer outcomes", "technology", "data-driven decisions", "data driven decisions", "decision making", "strategic planning", "execution", "delivery", "collaboration", "communication", "leadership", "mentoring", "negotiation", "presentation", "stakeholders", "roadmaps", "feature definition", "release management", "product launch", "market research", "voice of customer", } # Primary key — GLM 5.1 (most reliable, used for resume parsing + customization) NVIDIA_API_KEY = os.getenv("NVIDIA_API_KEY") GLM_BASE_URL = NVIDIA_BASE_URL GLM_MODEL = "z-ai/glm-5.1" # Centralized V1 model fallback chain (src/nim_fallback.py health-checks these in # order and selects the first healthy one). Edit here to change live V1 models. # As of 2026-08: glm-5.2 times out, mistral-small-4 is 410 EOL, nemotron is healthy. V1_MODEL_CHAIN = [ "nvidia/nemotron-3-super-120b-a12b", # verified healthy — placed first to avoid "z-ai/glm-5.2", # the ~60s probe cost of a hung model "mistralai/mistral-small-4-119b-2603", ] # ── 10 assessment models ──────────────────────────────────────────────────── # phase1 = used for quick keyword scoring (instant, no LLM needed now) # phase2 = used for detailed LLM assessment of top jobs ASSESSMENT_MODELS = [ # ───────────────────────────────────────────────────────────────────── # Phase 2 (detailed LLM assessment) speed results from test run: # Kimi-K2.6 ~5s/batch FAST ✅ phase2=True # Step-3.7-Flash ~8-35s FAST ✅ phase2=True # Qwen3.5-397b ~9s FAST ✅ phase2=True # Qwen3.5-122b-v2 ~12s FAST ✅ phase2=True # GPT-OSS-120b ~11s FAST ✅ phase2=True # Qwen3.5-122b ~40s OK ✅ phase2=True # DeepSeek-v4-Pro ~42s OK ✅ phase2=True # DeepSeek-v4-Flash ~229s SLOW ⚠️ phase2=False (too slow) # GLM-5.1 ~234s SLOW ⚠️ phase2=False (too slow) # MiniMax-M2.7 ~908s BLOCKED❌ phase2=False # ───────────────────────────────────────────────────────────────────── { "name": "Kimi-K2.6", "model": "moonshotai/kimi-k2.6", "api_key": os.getenv("NVIDIA_API_KEY_3"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~5s per batch FAST "tailor": True, # capable tier — strong at structured JSON rewrites }, { "name": "Step-3.7-Flash", "model": "stepfun-ai/step-3.7-flash", "api_key": os.getenv("NVIDIA_API_KEY_8"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~8-35s FAST "tailor": False, # weaker — drops roles / sparse bullets; not for tailoring }, { "name": "Qwen3.5-397b", "model": "qwen/qwen3.5-397b-a17b", "api_key": os.getenv("NVIDIA_API_KEY_7"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~9s FAST "tailor": True, # capable tier — large model, good instruction following }, { "name": "Qwen3.5-122b-v2", "model": "qwen/qwen3.5-122b-a10b", "api_key": os.getenv("NVIDIA_API_KEY_7"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~12s FAST "tailor": False, # smaller — keep for assessment, not tailoring }, { "name": "GPT-OSS-120b", "model": "openai/gpt-oss-120b", "api_key": os.getenv("NVIDIA_API_KEY_5"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~11s FAST "tailor": True, # capable tier }, { "name": "Qwen3.5-122b", "model": "qwen/qwen3.5-122b-a10b", "api_key": os.getenv("NVIDIA_API_KEY_4"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": True, # ~40s OK "tailor": False, # smaller }, { "name": "DeepSeek-v4-Pro", "model": "deepseek-ai/deepseek-v4-pro", "api_key": os.getenv("NVIDIA_API_KEY_2"), "base_url": NVIDIA_BASE_URL, "extra_body": {"chat_template_kwargs": {"thinking": False}}, "phase1": True, "phase2": True, # ~42s OK "tailor": True, # capable tier — strong reasoning }, { "name": "DeepSeek-v4-Flash", "model": "deepseek-ai/deepseek-v4-flash", "api_key": os.getenv("NVIDIA_API_KEY_6"), "base_url": NVIDIA_BASE_URL, "extra_body": {"chat_template_kwargs": {"thinking": True, "reasoning_effort": "high"}}, "phase1": True, "phase2": False, # ~229s too slow for phase2 }, { "name": "GLM-5.1", "model": "z-ai/glm-5.1", "api_key": os.getenv("NVIDIA_API_KEY"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": False, # ~234s too slow for phase2; used for resume customization }, { "name": "MiniMax-M2.7", "model": "minimaxai/minimax-m2.7", "api_key": os.getenv("NVIDIA_API_KEY_2"), "base_url": NVIDIA_BASE_URL, "extra_body": {}, "phase1": True, "phase2": False, # ~908s BLOCKED }, ] # ── Google Sheets / Drive ────────────────────────────────────────────────── GOOGLE = { "sheet_id": os.getenv("GOOGLE_SHEET_ID", "1Ehxt3eortehbtySdtgcSrMhCqmxIMUAmvRqSkII0HJk"), "sheet_tab": "Job Applications", "drive_folder": "Job Automation Agent — Resumes", "credentials_file": "google_credentials.json", # Service account JSON "token_file": "google_token.json", # OAuth token (auto-created) } # ── Job Search — Product Manager roles ONLY ─────────────────────────────── JOB_SEARCH = { "roles": [ "Product Manager", "Senior Product Manager", "AI Product Manager", "Technical Product Manager", "Product Owner", ], # Required words in job title — scraper DROPS any job without these "required_title_words": ["product"], # Blocked words in title — scraper DROPS any job with these "blocked_title_words": [ "teacher", "faculty", "engineer", "developer", "designer", "accountant", "analyst", "consultant", "manager data", "project manager", "program manager", "marketing manager", "sales", "hr ", "recruiter", "finance", "legal", "civil", "mechanical", "electrical", "hardware", "network", "software engineer", "data engineer", "ml engineer", ], "locations": ["India", "Bangalore", "Mumbai", "Delhi NCR", "Hyderabad", "Pune", "Remote"], "experience_min": 2, "experience_max": 15, "max_jobs_per_platform": 30, "days_posted": 7, "primary_role": "Product Manager", } # Platforms (Naukri blocked by Akamai) PLATFORMS = { "linkedin": True, "naukri": False, "indeed": True, "glassdoor": True, } # Assessment — process ALL PM jobs ASSESSMENT = { "min_score_for_llm_resume": 6, # LLM-tailored resume for score >= this "generate_all_resumes": True, # Template resume for ALL PM-relevant jobs "max_llm_resumes": 30, # Max LLM-tailored resumes per run "process_all_jobs": True, # Assess ALL scraped jobs (not just top N) "dedup_days": 30, # Skip jobs seen in last N days (0 = no dedup) "test_mode": False, # Set True for quick test run "test_jobs_limit": 10, # Max jobs in test mode } # ── Production LLM generation (provider abstraction + fallback chain) ────────── # The pipeline is model-independent: any provider may generate content, but # deterministic validation (schema + independent score + parse) decides if it # passes. Providers are tried in order until one yields a READY (>=90) resume. LLM_GENERATION = { "provider_order": ["kimi", "nvidia_primary", "deterministic"], "temperature": 0.1, "max_retries": 2, "require_json_schema": True, "fallback_on_schema_error": True, "fallback_on_low_independent_score": True, "fallback_on_parse_failure": True, } # ── Resume automation policy (AUTO_AGGRESSIVE) ───────────────────────────────── # Controls how aggressively plausible JD terms are auto-included and which # resumes are downloadable, so a 100-job batch runs mostly hands-off. AUTOMATION = { "automation_mode": "auto_aggressive", # What to do per risk severity (see candidate_fit.severity): "review_policy": { "LOW_RISK_AUTO_INCLUDED": "auto_include", "MEDIUM_RISK_REVIEW_RECOMMENDED": "auto_include_with_review_flag", "HIGH_RISK_NEEDS_CONFIRMATION": "ask_user", "BLOCKED_DO_NOT_INCLUDE": "exclude", }, # Which statuses are downloadable: "download_policy": { "READY_90_PLUS": "allow", "READY_90_PLUS_REVIEW_RECOMMENDED": "allow_with_warning", "WEAK_90_INTERNAL_ONLY": "block", "NEEDS_REPAIR": "block", "NEEDS_USER_INPUT": "block", "NOT_ELIGIBLE_LOW_FIT": "block", "PARSE_FAILED": "block", }, } # ── ever-jobs integration ───────────────────────────────────────────────────── from src.ever_jobs_bridge.platforms import INDIA_DEFAULT_PLATFORMS EVER_JOBS = { # Point this at a HOSTED ever-jobs sidecar to enable the 150+ extra # platforms on HF Spaces (where a local Node process can't run). Defaults # to the local dev sidecar. "api_url": os.getenv("EVER_JOBS_API_URL", "http://localhost:3001"), "default_platforms": INDIA_DEFAULT_PLATFORMS, "max_results": 25, "request_timeout": 30, "http_timeout": 120, "max_platforms_per_call": 20, } # ── Scraper stealth + anti-block (Scrapling) ───────────────────────────────── # The fetch layer (src/scrapers/fetch.py) uses Scrapling for real-Chrome TLS # impersonation and Cloudflare bypass, falling back to plain requests. These # settings are read mostly via env so they can be set as HF Space secrets. SCRAPER = { # Browser TLS fingerprint to impersonate for HTTP fetches. "impersonate": os.getenv("SCRAPER_IMPERSONATE", "chrome"), "timeout": int(os.getenv("SCRAPER_TIMEOUT", "20")), # Comma/newline-separated proxy URLs (http://user:pass@host:port). The ONLY # real fix for datacenter-IP blocking (e.g. LinkedIn/Indeed on HF). Rotated # round-robin. Empty = no proxy (best-effort fingerprint stealth only). "proxies": os.getenv("SCRAPER_PROXIES", "") or os.getenv("SCRAPER_PROXY", ""), # Use the Camoufox stealth browser for Cloudflare-protected sites. "use_browser_stealth": os.getenv("SCRAPER_BROWSER_STEALTH", "1") not in ("0", "false", "False"), } # Direct-company ATS boards (Greenhouse/Lever/Ashby public JSON APIs). These are # the most block-resistant bulk source on a shared cloud IP. Override/extend via # the COMPANY_ATS_BOARDS env var (JSON list) without touching code. If empty, the # scraper uses its built-in seed list (src/scrapers/company_ats.DEFAULT_BOARDS). def _load_company_ats_boards(): raw = os.getenv("COMPANY_ATS_BOARDS", "").strip() if not raw: return [] try: import json as _json boards = _json.loads(raw) return boards if isinstance(boards, list) else [] except Exception: return [] COMPANY_ATS_BOARDS = _load_company_ats_boards() # Output OUTPUT = { "excel_path": "data/output/reports/job_report.xlsx", "resumes_dir": "data/output/resumes/", "report_dir": "data/output/reports/", } # Resume RESUME = { "pdf_path": os.getenv("RESUME_PATH", "data/resume/resume.pdf"), }