A hat tip

Inspiration

Several people and teams have done valuable work helping researchers understand and draft the Tri-Agency narrative CV. Their public guides, documentation, and tools shaped how the community thinks about responsible research assessment, evidence over metrics, and telling a story of impact. This project is grateful for that groundwork.

Projects we admire

• tcvguide.ca — the University of Calgary Knowledge Impact team's interactive TCV Guide and AI assistant, including a "4-step method" (List, Define, Describe, Support) and a retrieval-augmented chatbot.
• Raad Fadaak's TCV Guide — the vision, consultation work, and technical execution behind that guide, including its editorial-partner framing of AI and DORA-aligned emphasis on qualitative evidence.
• University of Victoria Libraries — Narrative CVs — a clear library guide situating narrative CVs within responsible metrics, knowledge mobilization, and researcher-impact support.

Official Tri-Agency resources

These are the authoritative sources. When they differ from any third-party guide (including this site), follow the funding opportunity and the agency pages.

Hubs and overviews

• CIHR — Tri-agency CV — central hub with templates, webinars, applicant and reviewer resources.
• SSHRC — Tri-agency CV — overview, active opportunities, and forms links.
• NSERC — Tri-agency CV — overview and linked instructions, template, and FAQs.

Instructions, FAQ, and review

• Tri-agency CV instructions
• Tri-agency CV instructions (SSHRC)
• Tri-agency CV — Frequently asked questions (SSHRC)
• Tri-agency CV — Frequently asked questions (CIHR)
• Guidelines for reviewing the Tri-agency CV
• Tri-agency CV webinar — agency overview of why the format exists and how to navigate it.

Writing and assessing contributions

• CIHR — How to highlight your research contributions and impacts
• NSERC — Guidelines on the assessment of contributions to research, training and mentoring
• SSHRC — Forms and online application tools — includes the Word template and linked narrative-CV guidance.

University & library guides

Beyond the projects above, several Canadian institutions publish practical support for narrative and Tri-Agency CVs:

• University of Calgary — Knowledge to Impact / Narrative CVs — institutional framing of the TCV transition and links into the interactive guide.
• Western University — Narrative CVs — clear summary of the three TCV sections, page limits, and workshop materials for faculty.
• UVic Libraries — Narrative CVs — places the TCV in a broader research-impact and responsible-metrics context.

Responsible research assessment

The Tri-Agency CV sits inside a wider shift away from journal-based metrics toward evidence of quality, contribution, and impact:

• San Francisco Declaration on Research Assessment (DORA) — the international statement the Tri-Agencies have signed.
• CIHR — Responsible research assessment — how DORA shows up in CIHR policy and review practice.
• CIHR — DORA frequently asked questions — practical guidance on contributions beyond publications and metrics.

Local vs cloud LLMs

This assistant runs a small language model in the browser rather than sending CV text to a cloud API. The articles below explain that trade-off — privacy and offline use on one side; model size, download cost, and writing quality on the other.

Trade-offs and when to choose which

• web.dev — Benefits and limits of large language models — compares cloud providers with on-device LLMs: latency, offline access, ongoing cost, and keeping personally identifiable information off external servers.
• web.dev — Right-sized AI — argues for smaller, local, task-specific models when privacy, connectivity, or cost matter, and for cloud models when complex reasoning is required.
• web.dev — Choose your platform — a clear client-side vs server-side matrix (privacy, latency, capability, cost) and when hybrid designs make sense.
• SitePoint — The definitive guide to local-first AI — practical overview of browser and on-device stacks, with an explicit local-first pattern that only falls back to a server API with user consent.

In-browser engines (the stack this tool uses)

• MLC AI — WebLLM: a high-performance in-browser LLM inference engine — introduces WebLLM as private, GPU-accelerated, zero-setup inference with an OpenAI-style API.
• Ruihang Lai et al. — WebLLM (arXiv) — technical paper on retaining up to ~80% of native performance in the browser via WebGPU and compiled kernels.
• web.dev — Build a local and offline-capable chatbot with WebLLM — hands-on walkthrough of caching model weights and running chat entirely client-side.
• web.dev — Improve performance and UX for client-side AI — honest limits: large downloads, hardware variance, WebGPU checks, and keeping the UI responsive with workers.

Taken together, these sources support the design choice here: for sensitive CV and application text, local inference reduces disclosure risk; for frontier writing quality, cloud models still win — which is why this tool emphasizes evidence review over trusting the draft unedited.

Short literature review

A brief, non-exhaustive survey of recent peer-reviewed and preprint work in the two research areas this project touches: on-device language models, and narrative CVs in research assessment. It is provided for orientation, not as a systematic review.

On-device and small language models

A fast-growing body of survey work maps the move from cloud-only inference toward models that run on personal hardware. Xu and colleagues' comprehensive review of on-device language models frames the central tension as a balance between the resource limits of edge hardware and the benefits of reduced latency, data localization, and personalization, surveying compression techniques (quantization, pruning, knowledge distillation) and collaborative edge–cloud designs [1]. Complementary surveys of edge LLMs cover the full lifecycle from efficient model design to runtime inference, explicitly citing privacy and offline operation as motivations for local deployment [2], and catalogue efficient-inference strategies such as speculative decoding and model offloading for resource-constrained devices [3].

In parallel, "small language model" (SLM) research argues that models in the roughly 0.1–7B-parameter range — the class this tool uses — can be practical and reasonably capable while remaining private and affordable. Lu et al. benchmark dozens of open SLMs on capability and on-device runtime cost (latency, memory, energy), positioning them for edge intelligence [4], while Wang et al. survey SLM techniques, applications, and trustworthiness (including privacy and hallucination) in the era of LLMs [5]. The consistent finding across this literature: local/small models trade peak reasoning and generation quality for privacy, latency, and cost advantages, and are best paired with task scoping and human oversight — the exact posture this assistant takes.

Narrative CVs and research assessment

Empirical study of narrative CVs is still emerging. A review of opinion pieces and early experiments notes that, despite strong policy momentum (DORA and the reform-of-research-assessment agreements), there was little dedicated empirical research on the format, and synthesizes expert positions and pilot feedback instead [6]. More recent work begins to fill that gap: Kaltenbrunner and Woods, drawing on observation and interviews across two funding rounds of a Dutch Research Council programme, theorize narrative CVs as a way to "tame complexity" in peer review and report that they can stimulate richer discussion of what counts as excellence — while also provoking reviewer resistance and negative reactions to self-aggrandizing language [7][8]. This maps directly onto design decisions here: emphasize evidence and role over rhetoric, keep claims verifiable, and avoid inflated language that reviewers penalize.

1. Xu, J., et al. (2024). On-Device Language Models: A Comprehensive Review. arXiv:2409.00088. arxiv.org/abs/2409.00088
2. Zheng, et al. (2025). A Review on Edge Large Language Models: Design, Execution, and Applications. ACM Computing Surveys. doi.org/10.1145/3719664
3. (2025). Efficient Inference for Edge Large Language Models: A Survey. Tsinghua Science and Technology. doi.org/10.26599/tst.2025.9010166
4. Lu, Z., et al. (2024). Small Language Models: Survey, Measurements, and Insights. arXiv:2409.15790. arxiv.org/abs/2409.15790
5. A Comprehensive Survey of Small Language Models in the Era of Large Language Models: Techniques, Enhancements, Applications, Collaboration with LLMs, and Trustworthiness (2024). arXiv:2411.03350. arxiv.org/abs/2411.03350
6. Promoting narrative CVs to improve research evaluation? A review of opinion pieces and experiments (2023). Research Evaluation, rvad013. doi.org/10.1093/reseval/rvad013
7. Kaltenbrunner, W., & Buckley Woods, H. (2024). Taming complexity: narrative CVs in grant funding evaluations (RoRI Working Paper No. 14). doi.org/10.6084/m9.figshare.27291537.v1
8. Varga, J., & Kaltenbrunner, W. (2024). Narrative CVs: How do they change evaluation practices in peer review for grant funding? DORA blog. sfdora.org

Literature reviewed July 25, 2026. Citation details may vary by version; follow the linked source of record.

How this tool differs

This assistant is a distinct project with its own goals and design decisions:

• Local, not cloud. Instead of sending CV content to a hosted model API, this tool downloads a small model and runs inference in your browser via WebGPU. CV text, application context, and drafts are not intentionally sent to an external model service or this site's server.
• More pathways into your material. You can upload a CV, answer a guided interview, or paste notes and legacy text. All three routes converge on the same verified evidence model, which helps applicants whose most important work is poorly represented by a publication list.
• Evidence before prose. The workflow structures extraction, human verification, an evidence-gap audit, and discipline-aware importance ranking before any drafting, with explicit anti-invention rules and optional [EVIDENCE NEEDED] markers.
• Discipline-specific ranking. Candidate contributions are scored against signals tuned to SSHRC-, NSERC-, and CIHR-leaning contexts to help focus a section, rather than summarizing an entire CV.
• Honest trade-offs. The interface is candid that local models are less capable than leading cloud models, that "local" still downloads files, and that human verification is mandatory.

Original work

All code, content, prompts, discipline profiles, and ranking logic in this assistant are original and written independently for this project. No code was copied from the projects above; they are acknowledged here as inspiration only. Any resemblance is at the level of shared goals — helping researchers articulate impact under a narrative-CV model — not shared implementation.

Official agency pages remain the authority for templates, formatting, page limits, appendices, and opportunity-specific rules.

Acknowledgements and resource list reviewed July 25, 2026.