Technology and safeguards

PathwayAI is a career learning platform from InfraLoka. You bring your documents and a goal; it finds the skills between where you are and where you want to be, plans how to close them, and answers questions about your documents with page-level citations.

Organizations use the same foundation for structured coaching, an approved course catalog and a consent-first hiring pipeline. The rest of this page explains how each part works and where its limits are.

1. Document ingestion. Multi-page PDF resumes and policy manuals are parsed into page-addressable text, so every later answer can point at a page.
2. Embeddings. Passages and questions become 384-dimensional vectors from a multilingual sentence-embedding model, run locally on CPU. No text leaves the platform to be embedded.
3. Knowledge graph and database. Vectors are indexed in a self-hosted vector database with deterministic point ids and linked to a knowledge graph of skills, prerequisites, roles and courses.
4. Answers with citations. Roadmaps, skill-gap radar charts, audited PDF reports and chat answers are built from that evidence and cite the exact source page.

Vector size

384 dimensions, from the multilingual E5 small model.

Skill taxonomy

17 skills in the seed vocabulary, growing with the catalog.

Citation or refusal

Every chat answer cites its source or declines to answer. A test enforces this; it is not a sampled statistic.

Skills, roles, certifications and marketplace offers form one graph with prerequisite, covers, requires and related edges. The figures below are computed from the graph itself when you open it, not written by hand.

Two kinds of input

• PDF documents. Resumes, CVs and policy manuals, parsed page by page with the page numbers kept for citations (POST /api/documents, then extractPdfPages and chunkPages at 512 tokens), with OCR for scanned pages.
• Goals and constraints. A target role, a timeline, a weekly study budget and target companies, written in plain language and used to plan and replan the roadmap.

Four kinds of output

• Adaptive roadmaps. Phases with milestones, a progress checklist and a weekly study allocation.
• Skill-gap radars. Your current competencies against the target role's.
• Verifiable PDF reports. A saved roadmap exports as a PDF with an HMAC-signed QR code. Scanning it confirms PathwayAI issued the document, not that its contents are accurate.
• Grounded Q&A. A chat assistant that cites the exact document pages and clauses behind each answer.

Hiring on PathwayAI inverts the usual default. A learner's profile is not a product to be listed: every read of candidate data by someone other than the candidate passes through a disclosure grant the candidate created, naming the company, the role, the exact fields and an expiry. Revoking one takes effect on the recruiter's next read, not on a nightly job.

Open roles are public; no sign-in is needed to look. Browse them at /careers/jobs. You only sign in to see your own fit score or to apply.

How consent works

1. You opt in. Joining the candidate pool shows companies an anonymous match count only. You choose which fields can be shared at all; nothing identifies you until you apply.
2. You apply. Applying creates the grant and the application in one transaction, so a failed application never leaves a stray grant. The grant has an expiry from the moment it exists.
3. You can revoke. Every grant you have made stays listed, revoked ones included, with the date. Revoking marks you as consent-withdrawn on the recruiter's view immediately and withdraws your fit score with it.

Fit scoring

A fit score is computed from competencies extracted from your own resume by the same embedding model that powers the skill-gap radar. It never uses your name, gender, age, photo, address, nationality, institution, graduation year or employment gaps. That exclusion list is closed in code, and widening it requires re-running the bias tests before the change can merge.

• Null is not zero. Failing a required competency returns no score, with the gaps named, never a 0% bar. Each named gap links to the roadmap module that closes it.
• A score needs consent too. A score derived from your resume is a view of your resume, so recruiters see one only when you shared your competency profile. Recruiter-facing ranking also sits behind a kill switch; your own view does not.

Assessments

Interview questions adapt in difficulty to how you are doing, using bands a person assigned at approval time. This is deliberately not item-response theory: calibrated IRT needs response data this platform does not have, and claiming otherwise would make it an unvalidated psychometric instrument used as a hiring gate.

• Generated, then reviewed, then served. A model writes each question into a review queue, and a person approves it against a four-point checklist before any candidate sees it. No question is generated during a session.
• Practice leaves no trace. A practice run cannot be attached to an application and writes no result a recruiter could read. The service enforces this, not only the route.
• Your answers are data, never instructions. Answers reach the grader in a structured message channel, so text that looks like a command or a score is graded, not obeyed.

For companies

Hiring pipelines are directed graphs of stages, not fixed ladders: branches, parallel tracks and paths that rejoin are all possible. The server rejects cycles. Publishing a workflow freezes it; editing creates a new version, and applications already in progress keep the version they started on. Every read of candidate data is audited.

Personal secret vaults

A passphrase-derived key (PBKDF2 and AES-GCM, computed in your browser) gates the resume portal. The passphrase and key are never sent to or stored on the server. Documents are stored durably on the server, scoped to your tenant and account.

Encryption at rest

Uploaded documents, their search-index text and your quiz answers are encrypted at rest on the server with a server-held key. The vault passphrase only locks the resume portal view in your browser; it does not encrypt them, and it does not cover your question sets or flashcards.

Answers that cite their source

Recommendations and chat answers link back to the source document with page numbers, the supporting text and a confidence score.

Scanned documents

Scanned or image-only PDFs fall back to local OCR in English and Indonesian, so extraction works when a page has no selectable text.

Governed marketplace

Partner offers, syllabi and mentor slots are checked against real skill gaps, with affiliate disclosure and an audit log that cannot be edited.

Corporate coaching approvals

Corporate courses move from draft to submitted to admin review to approved; approval adds the course to the shared LMS question bank, and mentor fees go through a separate versioned proposal.

There are two pricing models and no subscription. Current rates and worked examples are on the pricing page.

Every new account starts with a free starter grant of credits. Credit top-ups create a real invoice through the payment gateway; credits are applied only after a signed webhook confirms the payment, so an unverified callback can never top up an account.

In the marketplace you can put several offers in a cart, get a quote computed on the server (subtotal, discounts, tax and fees as separate lines, with an expiry) and check out into one order per offer. The browser never sets the price.

Rahmatan lil ‘Alamin: a mercy and benefit for all creation.

Good career guidance, structured learning paths and fair hiring have mostly reached people with money, connections or the right school. PathwayAI exists so that a learner anywhere in Indonesia or Southeast Asia can plan their own path with the same quality of evidence.

1. Open access. Roadmap generation, resume skill extraction and document Q&A have no paywall.
2. Provenance. Every insight points to a source document, with page citations and a confidence score.
3. Your key, our storage. The resume portal key never leaves your browser; documents are stored on the server and readable only by your account within your tenant.
4. Honest commerce. No hidden affiliate links or pressure selling; marketplace partners must meet quality commitments.

• Built for learners in Jakarta, Bandung and other growing tech cities.
• Ongoing research on multilingual embeddings and graph-based retrieval.
• Documented schemas, open API endpoints and partner integrations.
• Regular bias audits and checks against answers the evidence does not support.