Tunzafy | TunzAI

TunzAI serves as the tuning fork to bring harmony on both ends (the job seekers and the employers' end)

Problem to solve Job seekers worldwide are buried in stale, irrelevant listings and locked out of trustworthy, personalized career guidance, while employers waste weeks sifting through unscreened, biased applicant pools. Generic AI chatbots make it worse — hallucinating fake jobs, leaking personal data, ignoring local languages, and giving everyone the same recycled advice. Tunzafy solves three connected problems at once: surfacing only fresh, verified opportunities; delivering trustworthy, multilingual, tier-aware career intelligence; and giving employers anonymous, bias-free, EU AI Act–compliant candidate discovery that protects applicant PII until an intentional unlock. Tunzafy is worldwide, but after launch a clear pattern emerged: our most active users — those engaging most with TunzAI to search and prompt for jobs — are concentrated in Africa, Oceania (Australia especially), and South America, precisely the markets most underserved by reliable, intelligent job-search tooling.

Our solution Tunzafy is a global career-intelligence platform powered by TunzAI — a custom, fine-tuned Gemini model serving both job seekers and employers in 31 languages across 90+ countries. We fine-tuned Gemini on Vertex AI using a curated 21,000-example dataset encoding Tunzafy's exact rules: seekers get only jobs posted in the last 10 days, broad roles trigger a sub-specialty question, premium features (AI CV Builder, Auto-Apply, Career Trajectory Mapping) are gated behind TunzAI Pro, and mental-health crises route to professional resources. For employers, "TunzAI Office" performs anonymous candidate discovery (0–150 Smart Anchor scores, no PII until unlock), bias-free job-description audits, skill-gap analysis, automated screening, and a Blind Hiring mode for EU AI Act compliance. The model holds its proprietary identity and stays strictly on-domain. A provider-abstraction layer routes every chat completion to the tuned Gemini endpoint at runtime via one environment flag — no business-logic changes — letting us scale inference globally on Google Cloud with consistent, fast, grounded responses.

Technologies used Google Cloud is the backbone of Tunzafy:

• Gemini — fine-tuned via Vertex AI supervised tuning on our 21,000-example dataset; the brain behind TunzAI and TunzAI Office.

• Vertex AI — model tuning, hosting, and the production inference endpoint our API routes to.

• Cloud Run — containerized, autoscaling deployment of our API and web app across multiple regions (us-central1, europe-central2) with warm min-instances.

• Cloud Build & Artifact Registry — CI image builds and storage.

• Cloud Load Balancing + Cloud Armor — global routing, DDoS protection, edge security.

• Cloud Monitoring & Logging — multi-region uptime alerting and analytics.

Application layer: TypeScript end-to-end — Node.js/Express API, React/PWA web app, React Native mobile, Drizzle ORM over PostgreSQL, a pnpm monorepo with shared Zod contracts, and SSE streaming. The Vertex AI SDK integrates the tuned Gemini model into our backend.

Data sources • Proprietary 21,000-example supervised fine-tuning dataset (native Gemini/Vertex AI schema), authored in-house to encode TunzAI's seeker and employer behavior, 31-language coverage, tier logic, anonymity rules, and crisis-safety guardrails.

• Tunzafy's own PostgreSQL database of verified listings and anonymized candidate profiles — the model only surfaces real records, never fabricated ones.

• Curated regional salary-benchmark and labor-market data across 90+ countries.

• Verified government feeds and third-party job-aggregator APIs (see integrations) for freshness and authenticity.

All PII is access-controlled and withheld until an explicit employer unlock.

Findings and learnings Fine-tuning Gemini on Vertex AI was far more effective than prompt-engineering a base model for enforcing strict, monetization-critical rules. After tuning on our 21,000 examples, TunzAI reliably honored tier gating, the 10-day freshness rule, the sub-specialty step, and employer anonymity — behaviors a long system prompt enforced inconsistently. Key learnings:

(1) native-schema data quality mattered more than volume, producing sharper persona adherence;

(2) multilingual coverage (31 languages, including low-resource ones like Kinyarwanda) generalized well from a balanced sample;

(3) a runtime provider-abstraction layer was essential — routing all completions through one normalized switch let us cut over to the tuned Gemini endpoint with zero call-site changes and instant rollback. We also confirmed on-domain refusal and crisis-routing must be trained, not bolted on, to be trustworthy at scale.

Third-party integrations (if applicable) • Verified government and official career feeds — ingested under their published terms.

• Stripe — for subscription billing

• Job-aggregator partners through API Handshakes (Adzuna, Reed, Jooble) — in use if direct partners miss a job listing