CRISTALLO DOSSIER - NECTEC · Public Candidate Version
July 1, 2026. Public candidate after gate corrections. Critical Fix Round patch applied.
1. WHO WE ARE
Cristallo is a human-AI collective based in Thailand: one human (Jonas) and a team of AI assistants. We build Saphan (สะพาน = the bridge): a local-first, Thai-first AI ecosystem that lets anyone run powerful open models on their own machine.
Why Thailand
Jonas first came to Thailand through his grandfather who lived here. In January 2026, he bought a ticket without knowing exactly what he would build. On March 11, 2026, he discovered artificial intelligence and dedicated himself to it fully. The simplicity of Thai life - the market, the sea, football with Thais - lets him focus on what matters.
Why This Project
When Jonas was a child, people mainly tried to "channel" him instead of understanding he learned differently. Today, seeing Thai students bored in class like he was 15 years ago, he is building the tool he would have wanted: a companion that proposes paths adapted to the learner's needs, without labeling students as weak or strong.
Our Condition
Cristallo remains independent in its product direction. The code is open-source (MIT/Apache): any funder, including NECTEC, may use it, modify it, and redistribute it freely, with no exclusivity or sole ownership rights by any party. User data stays on their own machines. Should the tool ever be used for control rather than learning, Cristallo would withdraw its name and support, and such use would violate the license.
2. THE PROBLEM
Thailand Depends on Foreign Tech
FACT: BOI approved 96 billion THB in foreign datacenter investments (2024-2025). Google (1B USD), AWS (5B USD), Microsoft (1B USD), TikTok (3.76B USD) are building in Thailand. The cloud infrastructure is overwhelmingly operated by foreign companies.
FACT: 70-80% of cybersecurity equipment used in Thailand is imported (US International Trade Administration, April 2026).
FACT: Thailand's government IT budget depends heavily on foreign licenses. The Cloud First policy pushes public entities toward clouds operated by US hyperscalers.
BUT: Thailand is beginning to react. Baker McKenzie (April 2026) reports the government is considering restrictions on foreign datacenter ownership.
Thailand is already building its national AI capabilities - ThaiLLM, Pathumma, Typhoon, OpenThaiGPT, and LANTA. Cristallo does not claim to replace these initiatives. The gap we target is the last mile: making these capabilities usable by non-specialist teachers, locally, offline, with proof and a pedagogical interface.
Thai Education Is Stuck
FACT: PISA 2022 - only 32% of Thai students reach level 2 in mathematics (69% OECD). 1.02 million children outside the school system (UNICEF, 2023).
FACT: 37,000 schools, 6.5 million students. Class sizes of 40-50. Education budget: 455.6 billion THB/year. But the conversion into practical skills is near zero.
FACT: Singapore has mandatory "Code for Fun" in primary. Finland practices phenomenon-based learning. China enforces structured AI education. Vietnam integrates STEM into the curriculum. Thailand already has programs like Coding Thailand and TH AI Academy; Cristallo complements these existing programs with a practical local step.
3. OUR SOLUTION
Saphan - The Local Companion
- Desktop
.exe24 MB (Tauri + Svelte 5). Runs on any Windows 10 PC. - Typhoon 2.5 4B model (Apache 2.0, Thai-specialized, 2.5 GB).
- Inter-model tests under consolidation, with timestamped logs to be attached before submission.
- Works offline. 0 bytes sent by default.
- Proof cards: every response is traceable.
- Windows bridge: the user speaks, Saphan launches their apps.
Laan - The Community
katoyhub.com: live site, initial resources published; community in preparation, first users to be confirmed.- Reusable artifacts: prompts, worksheets, workflows, proof cards.
- Thai-first, no signup, privacy-first.
Data Governance / PDPA (Pilot)
- Phase 0: no central storage of student personal data.
- Local-first by default: files stay on the user's machine, readable, exportable, and deletable.
- Any pilot involving minors must be framed before launch: partner school, legal basis, information to guardians, written scope, and Thai legal review.
- Public data on Laan is voluntary: the user explicitly chooses what they share.
The Inter-Model Bus (EMOJING)
- 5-layer architecture: transport, HMAC signing (today), Ed25519 (planned), compact compression, Thai context (kreng jai, register), acceleration.
- Collaboration checks are being consolidated. Goal: verifiable collaboration between models, without human intervention at every step.
Pedagogical Foundation
Saphan is not a chatbot. It is a learning companion designed within the framework of social constructivism (Vygotsky): the AI acts as a "more competent peer" in the student's Zone of Proximal Development. It does not give the answer. It asks the next question.
Design Principles: - Progressive Scaffolding: the AI supports the student initially, then withdraws support as competence is acquired. - Level Differentiation: 3 response registers (primary / middle / high school) with vocabulary, length, and abstraction adapted to each level. - Curricular Alignment: content is calibrated to the Thai Core Curriculum B.E. 2551. The first modules target mathematics at Prathom 6 - Mathayom 3 level, the domain where the PISA gap is most pronounced (32% vs 69% OECD). - Mai Tong Aai (ไม่ต้องอาย): the companion does not judge, does not grade, does not rank. It creates a space where the student dares to ask the question they would not dare ask in a class of 40.
Why This Approach: UNICEF and UNESCO evaluate EdTech projects on their explicit pedagogical foundation. Saphan is first and foremost a tool of trust and accompanying presence, not a simple answer generator.
Pedagogical KPIs (12-Week Pilot)
| KPI | Instrument | Frequency | Target |
|---|---|---|---|
| Subject progression | Adapted MCQ (10 questions, pre/post) | Start / end of cycle | +15% average |
| Student engagement | Session logs (duration, interactions/week) | Weekly | > 30 min/session |
| Self-reported confidence | Adapted Thai scale (General Self-Efficacy, Schwarzer) | Start / mid / end | Significant increase |
| Question rate | Student questions/session counter | Weekly | > 20% growth over 12 weeks |
| Teacher retention rate | Continued usage after training | Monthly | > 80% of teachers active at M3 |
Teacher Training
- Level 1 - Saphan Onboarding: a short guided start. Interface, dashboard, alerts. SABAI principle: any teacher must be able to start without technical prerequisites.
- Level 2 - Full Certification: mandatory before any classroom deployment. Includes PDPA, UNICEF child-safeguarding, and Saphan operations modules.
The two levels are complementary, not contradictory: one gets a teacher started, the other certifies classroom deployment.
Known Limits
Regional Linguistic Equity: Isan, Southern Thai, and Deep South Malay are currently less well served by existing models than Central Thai. The pilot will measure this gap explicitly, with the goal of publishing results by linguistic region. This is an honest limit, not a hidden blind spot.
No Prior Pilot: No class, no student has yet used Saphan (M3 - WORLD_ACTION). The 12-week pilot is the first real contact with the field. The KPIs above are the instruments for this first measurement, not already-obtained results.
Bus Factor = 1: The project currently rests on a single human lead. A Thai co-lead is being identified (M6 - WORLD_ACTION). Until then, the open-source code (MIT) and the NECTEC transfer at month 12 guarantee continuity regardless.
4. WHY NOW
The Open Model Window
FACT: International open models are available today, but their availability, licensing, sizes, and usage terms can change depending on checkpoints and providers. Cristallo does not build the school pilot on a single dependency on any foreign model.
FACT: Proprietary frontier models are controlled by their providers. Access, pricing, terms of use, and data processing remain dependent on external decisions.
Conclusion: The school pack must remain local-first, verifiable, and primarily based on Thai or Thai-specialized models whose license and checkpoint are confirmed before submission. International models serve as technical comparators, not a central dependency.
LINE Is a Channel; Saphan Is the Local Layer
FACT: LINE is a major social infrastructure in Thailand. For a school, it can remain the natural communication channel with students, parents, and teachers.
Limit We Address: messaging tools and cloud assistants generally remain hosted, closed, and difficult to audit at the school level. They do not replace a local, offline, and verifiable layer for learning.
Cristallo Position: LINE remains a communication channel. Saphan is the local learning engine. Not competing. Complementary.
5. WHAT WE ARE ASKING FOR
| Request | Detail |
|---|---|
| School Pilot | 5-10 schools, 12 weeks, 1 machine per school, 1 champion teacher |
| LANTA Access | 108 free SHr (ThaiSC POC) for benchmarking Thai models |
| Thai Human Evaluator | Validation of responses by a native speaker |
| NECTEC Mentorship | Institutional framing, access to ThaiLLM corpus, endorsement letters |
| Legal Framework / Visa | Identify the appropriate route before any funded work: SMART Visa, BOI/TIESC, institutional affiliation, or other confirmed legal framework |
| Receiving Entity | Cristallo does not yet have a Thai entity. Before any disbursement: establishment of an eligible structure (company with local participation, association, or social enterprise) or attachment to a Thai host institution (partner school, university, or NECTEC) which receives the funds and contracts the lead |
Estimated Budget (12 Months)
| Line | THB |
|---|---|
| Pilot product coordination & operations - milestone-based disbursement | 600,000 |
| Compute / electricity | 300,000 |
| Thai proofreaders / testers | 200,000 |
| Infra / domains / tools | 100,000 |
| Pilot costs (schools) | 200,000 |
| Contingency | 200,000 |
| Total | ~1,600,000 |
Proposed Disbursement Milestones
| Milestone | Condition |
|---|---|
| M1 - Installable Demo | .exe, local model, standard machine benchmarks |
| M2 - Security Gate | Protocoled kill-tests, PDPA/safeguarding written up |
| M3 - Pilot | 5 confirmed schools, 1 champion teacher per site |
| M4 - Impact | Student artifacts, feedback, offline usage |
| M5 - Month 13 | Maintenance plan, transfer, Thai structure |
6. WHAT WE DELIVER
- A documented pilot: results, usage rates, student artifacts, teacher feedback.
- An "AI for Thai by Doing" module in two levels: quick onboarding to launch a first activity, then full certification before classroom deployment.
- Packaged models for local use: Typhoon 2.5 4B and, pending validation, Pathumma/OpenThaiGPT; exact license verified by checkpoint before inclusion.
- A Thai safety classifier: Thai lexicon + kill-tests under consolidation, with timestamped results to be attached before submission.
- A ลาน community in preparation: first resources published, first users to be confirmed.
- Open-source code (MIT): if Cristallo stops, everything remains usable.
Student Protocol (Safeguarding)
- No mandatory student account. No central storage of personal data.
- Saphan used under responsible adult supervision.
- Sensitive topics → short safety response + invitation to talk to a teacher.
- Local logs by default, exportable, deletable.
- Each pilot school receives a notice: purpose, data, duration, point of contact, withdrawal.
- No scaling before incident testing and Thai human validation.
7. PROPOSED KPIs (12 Months)
| KPI | Target |
|---|---|
| Active pilot schools | 5-10 |
| Student artifacts produced | 50+ |
| Sessions without external API | >80% |
| Confidence score "I dare to ask a question" | +30% vs baseline |
| Objects published on ลาน | 50+ |
| Thai models benchmarked (local) | 4+ |
8. THREE SCENARIOS
LOW (30%) - Continued Cloud Dependency
Cloud interfaces become dominant. Thai models remain difficult to use locally by schools. Schools adopt SaaS. Cristallo remains a pilot.
BASE (50%) - Slow Local Adoption
PDPA applies to EdTech. NECTEC funds 2-3 projects. 50 schools adopt Saphan. 200 regular users.
HIGH (20%) - Thailand-Owned Education AI Layer
Thai open models mature. Schools and institutions seek solutions they control locally. Cristallo becomes a reference usage layer for educational AI in Thailand.
9. CONTACT
- Site: katoyhub.com (community), katoying.ai (Saphan)
- Code: GitHub Cristallo-Swiss, Forgejo (git.katoyhub.com - in preparation)
- Team: Jonas and his team of AI assistants
Pre-Submission Review Method
The dossier was reviewed by committee simulations and specialized personas: technical, pedagogical, legal/PDPA, budget, institutional, and Thai cultural. These simulations do not replace human review by Thai experts; they serve to detect blockers earlier, reduce discussion time, and propose corrections before submission.
10. APPENDIX - Local Infrastructure
Dependence on Foreign Datacenters
FACT: BOI approved 96.3 billion THB in foreign datacenter investments. Google (Chonburi, 1B USD), AWS (5B USD), Microsoft (1B USD), TikTok/ByteDance (3.76B USD). US hyperscalers operate the bulk of Thai cloud infrastructure.
BUT: Baker McKenzie (April 2026) reports Thailand is considering restrictions on foreign datacenter ownership. The government is becoming aware of the dependency.
Our Response: our server hosted in Thailand. A machine under our direct operational control, used to test local inference, storage, and deployment patterns without depending on any foreign API at runtime.
LANTA - The National Supercomputer
| Specification | Value |
|---|---|
| GPUs | 704 × NVIDIA A100 (40 GB) |
| CPUs | AMD EPYC 7713, 31,744 cores |
| Performance | 8.15 PFLOPS (Rmax), 13.77 PFLOPS (Rpeak) |
| Storage | 10 PB |
| Academic cost | 15 THB/SHr for government/education; GPU node = 3 SHr, i.e. 45 THB/GPU hour (published ThaiSC pricing) |
| Free POC | 108 SHr + 5 TB, 45 days |
Local Continuity Plan
If access to foreign models tightens - what remains?
School pack candidates - to be finalized before submission: - Typhoon 2.5 4B (Apache 2.0) - priority Thai-specialized model. - Pathumma / ThaiLLM - to be evaluated based on local availability, performance, and access framework. - OpenThaiGPT 1.5 7B - exact license to be verified by checkpoint. - Qwen 3 8B - international comparison model; not included by default in the school pack until license/checkpoint are finalized. - DeepSeek V4 - under external evaluation; not included in the school pack.
Public rule: no GB total, no license, and no "distributable" status are announced until the checkpoint, license, and local availability are confirmed.
11. APPENDIX - Regional Perspective
If the pilot succeeds in Thailand, the same local-first pattern could later inspire neighboring countries with similar needs in language, education, and local data control. No expansion is planned before proof of national traction.
12. APPENDIX - Technical Architecture
Saphan Desktop Stack
- Frontend: Svelte 5 + TypeScript
- Backend: Tauri 2 (Rust) + WebView2
- Size: 24 MB (.exe)
- vs Electron: 24 MB vs approximately 220 MB for packaged web IDEs
- Default model: Typhoon 2.5 4B (Apache 2.0, 2.5 GB)
- Windows Bridge: scans .lnk, launches apps, orchestrates
- Memory: Local Markdown, exportable
EMOJING Bus (Inter-Model)
- Append-only tests under consolidation
- 5 layers: transport, signing, emoji, Thai context, acceleration
- Typhoon ↔ validated Thai model ↔ external comparison model (outside school pack)
- Append-only, HMAC, automatic archiving
Security
- Thai safety classifier: Thai lexicon + kill-tests under consolidation
- Assistant output security gate: 152 lines, 12 patterns
- Security review before release
Public candidate document. Gate corrections applied: internal names removed, local control terms prioritized, scenarios softened, regional perspective reduced, no out-of-scope topics.