SkillBridge PH: A 10-minute voice interview in Filipino that turns your experience into a TESDA path and a local job

Problem Statement

The IT-BPM sector employs roughly 1.7 million Filipinos, and generative AI is automating the voice and chat support tasks it was built on. Many displaced agents, along with informal workers, returning OFWs, and out-of-school youth, can’t describe their transferable skills in the formal, English-heavy terms that job portals and training catalogs use. Public Employment Service Offices (PESOs) in LGUs are understaffed. TESDA runs hundreds of qualifications, but workers rarely know which ones fit their experience or where they lead. As a result, people re-skill blindly, drop out of courses, or slide into informal work while employers still report skills gaps.

Proposed Solution

SkillBridge PH is a voice-first web and mobile app, and also a phone line, that runs a friendly, adaptive “kwentuhan” interview in Filipino, Taglish, Cebuano, or Ilocano. An LLM asks follow-ups such as “Ano ang ginagawa mo kapag galit ang customer?” and maps the answers to a structured skills profile based on PSOC occupations and TESDA competency units. The system then ranks three things: (1) TESDA qualifications and short courses that close the gap, (2) free upskilling options such as TESDA Online Program modules, and (3) local openings from PhilJobNet and employer partners. Each comes with a plain-language explanation of why it was suggested. The worker gets a downloadable “Skills Passport” and a resume draft. With consent, the PESO officer gets a dashboard of candidates by barangay, skill cluster, and readiness, so job fairs and training slots can be targeted.

Target Market

• Primary users: displaced or at-risk BPO agents, informal workers, returning OFWs, and senior high and college graduates, potentially millions of jobseekers each year.
• B2G buyers: LGU PESOs (around 1,600 cities and municipalities), DOLE regional offices, TESDA provincial offices, DICT, and OWWA reintegration programs.
• B2B buyers: BPO firms running redeployment and outplacement programs, IBPAP, and large employers hiring for skills rather than diplomas.

How AI Is Used

• Speech-to-text and text-to-speech (Filipino and regional languages): lets low-literacy or non-English-confident workers take part. Without it, the tool becomes another English form.
• LLM adaptive interviewer: asks dynamic follow-ups to surface tacit skills like de-escalation, multitasking, and data entry speed. A static questionnaire can’t probe like this.
• Skill extraction and taxonomy mapping (LLM plus embeddings): turns free-form stories into PSOC and TESDA competency codes. This is the core matching step and can’t be done reliably with keyword rules.
• Semantic matching and ranking: links profiles to courses and job posts by meaning, not exact titles.
• Explainable recommendations: the LLM writes a “bakit ito” rationale for each suggestion, which builds trust and helps PESO counselors.

PLDT / Smart / ePLDT / Rokid Angle

• Smart: zero-rated or bundled data access, plus SMS and USSD fallback for interview reminders and job alerts.
• ePLDT cloud and data centers: host the profiles locally to meet Data Privacy Act data residency expectations. This is a natural B2G SaaS sale to LGUs and DOLE.
• PLDT Enterprise: sell it as a workforce-transition package to BPO clients that already buy PLDT connectivity and contact-center solutions.
• Rokid (stretch): AR-guided skills checks for hands-on TESDA trades, such as viewing a wiring task through the glasses and getting AI feedback.

Demo Scenario (3 minutes for judges)

1. Introduce “Jessa,” 29, a Pampanga call-center agent whose account was just replaced by an AI chatbot.
2. Live, Jessa speaks Taglish into the app. The AI asks three adaptive follow-ups about her handling of irate customers and her Excel reporting.
3. Her skills profile appears on screen: customer de-escalation, data entry, and basic troubleshooting, each mapped to competency codes.
4. The recommendations show TESDA Contact Center Services NC II upgrade to Bookkeeping NC III, a free TESDA Online module, and two local openings (a healthcare virtual assistant and a bank back office role), each with a “bakit ito” explanation.
5. Jessa receives an SMS with her Skills Passport link.
6. Switch to the PESO dashboard, where the San Fernando cluster shows 40 similar profiles. The officer schedules a targeted job fair.
7. Close: “Hindi ka napalitan. Na-upgrade ka.” (You weren’t replaced. You were upgraded.)

30-Hour Build Plan

Suggested stack: Next.js or Flutter web, FastAPI, Whisper (large-v3) or Google Speech-to-Text for Filipino, Gemini or GPT-class LLM with JSON output, multilingual-e5 embeddings, pgvector or Chroma, Twilio or Semaphore SMS gateway (Smart network in production). Public data / APIs: TESDA Training Regulations and qualification registry, TESDA Online Program catalog, PSA PSOC occupation codes, PSA OpenSTAT Labor Force Survey, PhilJobNet public listings, data.gov.ph, Hugging Face Filipino speech datasets (e.g., FLEURS fil_ph).

Risks & Privacy Notes

• ASR errors on regional accents and code-switching: mitigate by showing the transcript for the user to confirm and by having the LLM re-ask when it isn’t confident.
• Biased or wrong recommendations: position outputs as advice reviewed by PESO counselors, never as automatic rejections. Audit results across gender, age, and region.
• Data Privacy Act 2012: collect explicit, granular consent, especially before sharing with a PESO or employer. Minimize data by storing extracted skills, not raw audio. Process locally in a PH data center. Support the rights to access and erasure.
• Clause 24 compliance: the hackathon demo uses synthetic personas only. No real worker data is fed into AI tools, and all AI-generated components are disclosed.

AI Use Disclosure (draft)

We use Whisper or Google Speech-to-Text for Filipino transcription, a Gemini or GPT-class LLM for the adaptive interview, skill extraction, and recommendation explanations, and multilingual-e5 embeddings for matching. GitHub Copilot or similar tools helped with boilerplate code, and an LLM helped draft parts of this proposal. All personas and sample interviews are synthetic. The team reviewed and edited all AI outputs, taxonomy mappings, and final recommendations.

Rubric Scoring