Curtis Forbes, PhD — AI Evaluation و Educational Technology
Curtis Forbes هو باحث ومطور مقره في Edmonton، حاصل على PhD في تاريخ وفلسفة العلوم والتكنولوجيا من University of Toronto. يقوم بتطوير أدوات AI evaluation و educational technology، بما في ذلك Edmonton Daily Brief و Champollion، وهي بنية تحتية مفتوحة لـ low-resource machine translation.
تشمل المشاريع النشطة Game Day Suits (ملابس فاخرة لأيام مباريات الهوكي)، و Siphraya Bespoke (الخياطة التفصيلية)، والاستشارات وتوريد التصنيع الآسيوي من خلال SBMA Freeport Zone، وتطوير البرمجيات المخصصة لـ SMEs.
يركز العمل الأكاديمي على الواقعية العلمية والتحول البراغماتي في فلسفة العلوم، مع منشورات في Synthese وMetascience، وعمل تحريري لـ Spontaneous Generations.
تشمل مشاريع البرمجيات Edmonton Daily Brief (ملخص يومي آلي للسجل العام في Edmonton)، ومحاكاة فيزيائية تفاعلية، وتركيبات فنية تعتمد على AI، وبنية تحتية مفتوحة للترجمة للغات low-resource (Champollion)، وتطبيقات أعمال خاصة للعملاء في القطاعين الصناعي والزراعي.
curtis forbes
إنجاز المهام بمنطق.
FORBES BIOS 5.0 - جوهر هيولومورفي
hylomorphOS 5.1.0 · build 2026.10.07-88c1b0e
فحص الذاكرة: 12 عامًا من التصنيع · 7 منشورات · 30 لغة محلية - موافق
تحميل /home/curtis ... تم
جارٍ تحميل اللغات المحلية (30) [████████████] OK
جاهز. انقر فوق ملف - أو اكتب أمرًا. 'مساعدة' تساعد.
curtis@forbes:~$ tree
~/
│
── README.md
Curtis Forbes — philosopher of science turned manufacturing executive turned AI developer.
PhD, History & Philosophy of Science (Toronto, 2017). President, International Recreation Corporation (until 2025). Currently building epistemic tooling.
form/ is the research. matter/ is the business. Click any file, or type at the prompt — `help` works, and `ls` doesn't show everything ... ;)
│ │
── form/hps.md
Research area: the scientific realism debate — pragmatic, existentialist, and decision-theoretic approaches. Full record: Google Scholar.
## Education
- PhD, History & Philosophy of Science — University of Toronto, 2017. Supervisor: James Robert Brown. SSHRC Doctoral Fellowship; Bombardier CGS; Ontario Graduate Scholarship.
- MA, Philosophy — University of Victoria, 2008. Kenneth Rankin Prize for Best MA Thesis.
- BA with Distinction, Philosophy — University of Alberta, 2006.
## Publications
- A Decision Theoretic Approach to the Scientific Realism Debate — Synthese, under review, 2026
- The Pragmatic Turn in the Scientific Realism Debate — Synthese, 2024 (with Sandy Boucher)
- [Editor's Introduction: The Future of the Scientific Realism Debate] (https://spontaneousgenerations.library.utoronto.ca/index.php/SpontaneousGenerations/issue/view/1954) — Spontaneous Generations, 2018
- A Pragmatic, Existentialist Approach to the Scientific Realism Debate — Synthese, 2017
- Toward a Philosophy of Commercialized Science — review essay on Radder, The Commodification of Academic Research — Metascience 22(3), 2013
- Editor's Introduction: Science and Public Controversy — Spontaneous Generations, 2011
- Review: van Fraassen, *Scientific Representation* — Spontaneous Generations 3(1), 2009 · doi:10.4245/sponge.v3i1.6666
Dissertation: A Pragmatic, Existentialist Approach to the Scientific Realism Debate — University of Toronto, 2017.
## Teaching
Course instructor, University of Toronto, 2011–2016: Scientific Revolutions I & II, Topics in Philosophy of Science, Revolutions in Science. Teaching assistant, seven further courses, 2008–2016.
## Selected talks
Keynote, Bern 2024 · Hobart 2020 · Regina 2018 · Dubrovnik 2016 · Indianapolis 2016 · Toronto 2015 · Cape Town 2014
## Service
Referee: Philosophy of Science, Erkenntnis, Synthese, International Studies in the Philosophy of Science, Journal of Philosophical Research, Science & Education. Managing editor, Spontaneous Generations, 2010 and 2018.
│ │
── form/comp-ling.md
Machine translation for the languages that commercial providers forget.
Interests: MT evaluation and metrics · language preservation and revitalization · low resource and morphologically rich languages · neural MT · rule based MT · hybrid architectures · novel tokenizer architectures
CCHRF++: Linguistically corrected CHRF++ for cross-lingual quality comparison, based on relative character counts and variety.
MT-Eval Harness: Method-neutral machine translation evaluation harness with automated dynamic corpora and multiple metric loading, with authenticated CLI publication to MT-Eval Arena (Python).
MT-Eval Arena: Secure database for recording published evaluation runs with scoring, significance, and contest/prize mechanics. Includes integrity/sovereignty machinery such as DB-triggered quarantine gates, provenance rules, authorization backbone, and curl-to-bash benchmark queue runner (Postgres).
Champollion CLI: Method-neutral machine translation localization tool with custom routing mechanism that optimizes based on arena data.Includes translation memory and quality gate and supports all major translation methods (Node.js/JavaScript).
Shared Registries: Cross-runtime SSOT schema for methods, metrics, models, and licenses (JSON).
LYSS plug-in: Language-specific automated error detection, classification, and correction standard (Python).
CRK-translate: Hybridized English-to-Nêhiyawêwin (Plains Cree) translation method using Alberta Language Technology Laboratory's finite state transducer. Community consultation planned before any release (Python).
Champollion Atlas: Expansive typological classification of 7,000+ human languages from diff use open-source resources, plus a language documentation resource index (Postgres + JSON/SQLite distribution).
Champollion website: Docusaurus 3 website with animated project visualization (React/JavaScript).
Champollion MCP server: Exposes the benchmark queue, language search, cost estimation, and the translate pipeline (with translation memory and quality gate) to AI agents over the Model Context Protocol (Node.js/JavaScript).
Champollion Shared Task Sandbox: Air-gapped method execution infrastructure for community-hosted and -maintained machine translation contests. Prevents benchmark data contamination and ensures community data sovereignty simultaneously (Python/Docker).
Champollion Translate API (in development): Metered, IP-protected convenience layer for routing through multiple proprietary translation methods, hosted server-side to protect sovereign IP (Node.js/Postgres).
│ │
── form/benchmarks-evals.md
Most corporate and commercial AI benchmarks bias towards saleable properties: coding, tool use, recall, etc. This creates an incentive for developers to "train to the benchmarks," leading to benchmark decay through eval contamination. The epistemic question, asked of machines: when is a benchmark reliable?
Two answers:
1) A benchmark is reliable if not only its answers but also its questions are secret, and only the results are published
Example: Champollion — MT evaluation using air-gapped sovereign benchmarks with public evaluation rankings, with contamination lanes to keep memorized test sets from masquerading as quality.
2) A benchmark can be reliable if its correct answers are unknown at test time
Example: GDS AI Playoff Draft · GitHub — a deautomated, non-trainable benchmark: frontier models compete as general managers in NHL fantasy drafts. Tests multi-agent reasoning, competitive strategizing, budget discipline, rule-following, and decision-making under constraints; deliberately slow, world-adjudicated, and highly resistant to being trained against.
│ │
── form/data-sovereignty.md
Working position: language data belongs to the people whose language it is. Systems should be designed so that ownership is structural, not promissory.
- Sovereign benchmarks (Champollion) — test sets that communities create, own, and control, and that the platform never sees. Scores flow out; data never does. Built with professionals and speakers, with licensing enforced in code.
- Public access — (Prior Land) - data amalgamation software for public pollution and ground contamination data sets, build for retail property assessment (below Phase 1 impact assessment).
- BOYK - (Cosmic Consultant) - bring your own keys options for custom Bazi consultant, built with RAG and deterministic calculators.
│ │
── form/ai-education.md
Working position: the point of AI education is operator independence — people who can build, run, and govern their own systems. Build it, make it operable, hand it off.
- The simple-builder approach — deterministic tools, logic-as-data, provenance you can read: systems simple enough for their owners to maintain without the builder. Adoption dies when the builder walks out and nobody internal understands what the system does or how to govern it.
- Practiced, not preached — the AFI Quote Machine methodology was adopted and is run internally by the client. The same principle drives the tutoring practice: build the machinery of learning, not the product.
- Right to turn off — dependence on a vendor for your data and dependence on a builder for your capability are the same governance failure. Users deserve control of their entire stack.
│ │
── form/rotary.md
- Rotary Club of Manila Expats (RCME) · Charter Member. Organizing the club's speaker series and running charitable missions across the region — mission planning and logistics, community development initiatives.
Site: rotarymanilaexpats.com
│ │ │ │
── form/epistemic-toolworks/champollion/README.md
- Champollion · open-source, researcher-grade infrastructure for developers building machine translation methods for low-resource languages. Launching Q4 2026. champollion.dev
│ │ │ │
── form/epistemic-toolworks/champollion/cli.md
An internationalization CLI: `npx champollion sync` auto-detects your locale files, their formats, and the target languages — translates what's missing, skips what's done.
- Pre-commit ready — built to sit in a husky/lint-staged hook with SHA-256 change detection, so only edited strings re-translate. This site runs it exactly that way on every commit.
- Bulk & structured translation — JSON, TOML, YAML, Hugo Markdown (frontmatter + body), and XLIFF interchange documents (bulk XML), with format auto-detection.
- Method-neutral — the translation method is configurable per language pair: commercial APIs (Google Translate, DeepL), LLM backends (OpenRouter's 200+ models, OpenAI, Anthropic, Gemini), and community-developed plugins benchmarked in the MT Eval Arena.
- Quality gates — wrong-script detection, length-inflation checks, source-echo rejection; intelligent batching (per-backend batch sizing); guards against prototype pollution, path traversal, and code-block corruption in Markdown.
Source-available on npm as `champollion` (PolyForm Noncommercial). Formerly i18n-rosetta; this site's thirty locales run on it.
│ │ │ │
── form/epistemic-toolworks/champollion/arena.md
The evaluation half: a harness and leaderboard for comparing MT methods across shared corpora — scoring, hashing, and surfacing results. Contamination lanes keep memorized test sets from masquerading as quality; a one-line contribution path lets the crowd donate compute rather than one local run vanishing when the terminal closes.
- Licensing — AGPL-3.0, with an eval-standard plugin exception permitting separately-licensed eval plugins via the public plugin interface.
- champollion-lyss — the nêhiyawêwin eval-standard plugin: FST/gloss validation, fail-honest, no bundled data; use by written permission under Cree data-sovereignty terms (comp-ling).
Research context: benchmarks-evals. Source: gds-mt-eval-harness, public.
│ │ │ │
── form/epistemic-toolworks/champollion/harness.md
- MT-Eval Harness · Method-neutral machine translation evaluation harness with automated dynamic corpora and multiple metric loading, with authenticated CLI publication to MT-Eval Arena. Python.
│ │ │ │
── form/epistemic-toolworks/champollion/atlas.md
- Champollion Atlas · Expansive typological classification of 7,000+ human languages from diverse open-source resources, plus a language documentation resource index. Postgres, with JSON/SQLite distribution.
│ │ │ │
── form/epistemic-toolworks/champollion/registries.md
- Shared Registries · Cross-runtime SSOT schema for methods, metrics, models, and licenses. JSON.
│ │ │ │
── form/epistemic-toolworks/champollion/lyss.md
- LYSS plug-in · Language-specific automated error detection, classification, and correction standard. Python.
│ │ │ │
── form/epistemic-toolworks/champollion/crk-translate.md
- CRK-translate · Hybridized English-to-Nêhiyawêwin (Plains Cree) translation method using the Alberta Language Technology Laboratory's finite state transducer. Community consultation planned before any release. Python.
│ │ │ │
── form/epistemic-toolworks/champollion/mcp-server.md
- Champollion MCP server · Exposes the benchmark queue, language search, cost estimation, and the translate pipeline — with translation memory and quality gate — to AI agents over the Model Context Protocol. Node.js/JavaScript.
│ │ │ │
── form/epistemic-toolworks/champollion/sandbox.md
- Champollion Shared Task Sandbox · Air-gapped method execution infrastructure for community-hosted and -maintained machine translation contests. Prevents benchmark data contamination and ensures community data sovereignty simultaneously. Python/Docker.
│ │ │ │
── form/epistemic-toolworks/champollion/translate-api.md
- Champollion Translate API (in development) · Metered, IP-protected convenience layer for routing through multiple proprietary translation methods, hosted server-side to protect sovereign IP. Node.js/Postgres.
│ │ │
── form/epistemic-toolworks/champollion/website.md
- Champollion website · Docusaurus 3 website with animated project visualization. React/JavaScript. champollion.dev
│ │ │
── form/epistemic-toolworks/daily-brief.md
- Edmonton Daily Brief · An automated daily brief of the Edmonton public record, drawn from multiple public data streams (ER wait times, ETS tracking, weather, council voting records, etc.). Published automatically, with four verification gates in place of an editor: verbatim provenance, numeric and nominal fidelity, cross-vendor entailment checking, and legal adjudication for court items. Every item is deterministically cited to its source, and the system is designed so that the worst failure is an omission rather than a falsehood.
Site: edmontondailybrief.ca
│ │ │
── form/epistemic-toolworks/prior-land.md
- Prior Land · Aggregation and visualization of public real-estate data for realtors, buyers, and home inspectors. In development.
│ │
── form/epistemic-toolworks/cosmic-consultant.md
- Cosmic Consultant · Deterministic Bazi calculator with experimental data visualization and RAG-based LLM consultant. Private access only.
│ │
── form/art/dixie-flatline.md
- Dixie Flatline · dixieflatline.ai · 2026 — voice-first AI construct: custom persona inference, a cloned voice, React Three Fiber volumetric frontend, persona persistence. A ROM of the construct is slotted into this terminal: `cat .dixie.rom`.
This gallery has a lower floor. The mantas were never really retired — try `manta` at the prompt.
│ │
── form/art/bertagpt.md
- BertaGPT · bertagpt.com · 2026 — a late-2022 chatbot clone where the assistant is Berta: a headbanger from Beverly with a 2008 F-150, a denim vest, and total confidence about things he half-knows. A joke site whose engineering problem is character durability under adversarial input — server-built prompts, signed transcripts, drift reminders. 18+.
│
── form/art/ai-gm-league.md
- AI GM League · draft tracker · 2026 — thirteen frontier AI models and a control bot run NHL fantasy teams for a full season, scored daily from official NHL stats. Live standings, rosters, trades and trash talk. The successor to the GDS AI Playoff Draft.
│ │
── matter/game-day-suits.md
- Game Day Suits Ltd. · Founder, sole owner & President · 2025–present
Canadian importer of custom menswear through partnerships with siphrayabespoke.com, my tailoring operations in Bangkok, and damalerio.com, my fashion studio in Olongapo City Philippines.
Corporate vehicle for the enterprise consultancy and AI development practice.
Store: gamedaysuits.ca
│ │
── matter/manufacturing.md
- President — International Recreation Corporation (IREC), a subsidiary of International Play Company (iPlayCo) · Subic Bay · 2022–2025
Commercial play-equipment design, manufacturing, and installation for 20+ countries. 130+ employees, nine departments, 5,000+ SKU library. COVID-19 restructuring: 130 → 30 staff with manufacturing capacity preserved, profitability improved, all international delivery obligations met.
- General Manager — IREC · 2019–2022
Returned post-acquisition to restructure operations; maintained production through pandemic-era supply-chain disruption and sustained lockdowns.
- Co-founder, Director of Operations, CFO — IREC · Subic Bay Freeport · 2013–2016
Built the company from startup: 7 → 40+ employees in under three years on minimal capital. Acquired by International Play Company (iPlayCo, TSX-V) in 2016.
## Services
Manufacturing sourcing (fiberglass, carbon fibre, steelwork, upholstery, carpentry, sewing, injection moulding) · SBMA startup consulting · corporate services via stagecoach · interim management
│ │
── matter/siphraya-bespoke.md
- Siphraya Bespoke · Partner
Bespoke tailoring in Bangkok — guild-tradition craftsmanship with modern logistics. Manufacturing partner for the Game Day Suits menswear line.
Site: siphrayabespoke.com
│ │
── matter/damalerio.md
- Damalerio · Fashion studio · Olongapo City, Philippines
Pattern making, sample development, and garment manufacturing at small scale, with a designer and specialty-manufacturer network through the Fashion Institute of the Philippines. Production studio for Game Day Suits.
Site: damalerio.com
│ │
── matter/jewelry-empire.md
- Jewelry Empire · Co-founder · 2021–present
Jewelry retail through live-selling sales channels — bulk 18K gold fashion jewelry moved direct to buyers. US$2.5M+ annual revenue on ~US$100K initial investment; operated day-to-day by my equity partner, growing and paying dividends.
Retail and sales channeling, not manufacturing.
│ │
── matter/stagecoach.md
- Stagecoach · Manila, Philippines
Executive transport and logistics, plus corporate services in the Subic Bay Freeport Zone. Administrative and compliance support for Philippine operations.
Site: stagecoach.ph
│ │
── matter/software.md
AI-assisted software development and enterprise consultancy, via Game Day Suits Ltd. The throughline: deterministic systems the client's own people can own and operate. Reliability, control, and data sovereignty always prioritized.
Sample work:
- AFI Quote Machine — Deterministic industrial-equipment configurator for Australian Farm Innovations: provenance-traced calculation engine, logic-as-data architecture, integrated AI guidance. Next.js, React, Supabase. Client adopted the methodology and deployed internally.
- Fairfolk Farm — crop-share booking system for a community farm. In development.
- Prior Land — Aggregation and visualization of public real-estate data for realtors, buyers, and home inspectors. In development.
- Cosmic Consultant — Deterministic Bazi calculator with experimental data visualization and RAG-based LLM consultant. Private access only.
- Dixie Flatline — voice-first AI construct: custom persona inference, a cloned voice, React Three Fiber volumetric frontend, persona persistence. dixieflatline.ai
- Champollion — Machine Translation developer ecosystem with open-source, research-grade tooling. Launching Q4 2026. champollion.dev
│
── matter/tutoring.md
- Private tutor — Tutors International · 2017–2019
Specialist "super tutor" placement for a post-secondary student: full-time, live-in. Diagnostic assessment, individualized curriculum design, adaptive pedagogy. Contract renewed for a second year after transformative academic improvement.
Focus: students from globally mobile families — capable and ambitious, with hidden gaps in executive function, study methodology, and subject foundations left by discontinuous schooling. The work builds the systems, not just the grades.
## Services
- Hourly consultation — subject mastery, study-system design, assignment strategy, exam preparation.
- Placement & traveling tutor — full-time immersive support: residential or traveling, daily academic structure, executive-function coaching, university liaison.
── contact.md
- WhatsApp — +1 (825) 963-7848
- Email — gamedaysuits@gmail.com
Edmonton, Alberta. Always appreciate an excuse to go to Paris.
# انقر على مسار — أو اكتب أمرًا. جرب 'help'.