A buildless, dependency-free portfolio for GitHub Pages. The home page introduces the research; dedicated pages document two research studies, eight experience entries, and eight projects. All content and navigation work without JavaScript. JavaScript only enhances the project filters.
python3 -m http.server 8765 --bind 127.0.0.1Open http://127.0.0.1:8765. GitHub Pages serves the checked-in HTML directly; no server-side application or package installation is required.
- Edit experience and project content in
data/portfolio.json. - Edit research accounts, the home page, and shared templates in
scripts/build_site.py. - Edit styles in
css/portfolio.cssand filtering injs/portfolio.js. - Run
python3 scripts/build_site.pyand include the generated HTML in the same commit. - Run
python3 scripts/check_site.pyto check local links, fragments, images, and page metadata. - With the preview server running, run
python3 scripts/check_http.pyto request every unique internal and external HTTP destination, follow redirects, and verify PDF MIME types. LinkedIn blocks automated requests; its profile CTA was separately checked in a signed-in browser.
The earlier animation files remain in the repository for history, but the rebuilt site does not load them, third-party scripts, or external web fonts.
Two documents serve different purposes:
Resume/Aryan Dua - Resume.pdf: a designed one-page application resume.Resume/Aryan_Dua_Detailed_CV.pdf: an eight-page master CV with research methods/results, every experience and project, skills, and distinctions.
Editable HTML companions sit beside the PDFs. Shared typography and layout live in Resume/documents.css; document content and page composition live in scripts/build_documents.py, drawing project metadata from data/portfolio.json.
Rebuild both HTML sources and export both PDFs with:
python3 scripts/build_documents.py
python3 scripts/export_documents.pyThe export guards the expected one/eight-page counts and updates the legacy Resume.pdf URL only after both checks pass. Review rendered pages after editing: pagination checks alone do not establish visual quality.
For a quick export of only the existing resume HTML:
python3 scripts/export_resume.pyThis uses local Chrome/Chromium and updates both Resume/Aryan Dua - Resume.pdf and the legacy Resume.pdf URL. Poppler's pdfinfo is required: the exporter refuses to update the legacy PDF if the document is not exactly one page. Browser printing should use A4, CSS-defined margins, and no browser headers or footers.
The research accounts were prepared from the complete supplied Resume/Thesis.pdf and Resume/Asynchronous_Projective_Splitting.pdf. The complete controversy and WebNLG reports were also read, including result tables, limitations, and the WebNLG contribution statement. Their hosted PDFs and supplied Google Drive copies are linked from their project accounts and CV pages. Experience and project content uses the supplied resume HTML/PDF and the previous website. Degree naming is B.Tech + M.Tech, explicitly confirmed by Aryan.
- The Wipro / IIT Delhi internship and master's thesis are presented as one research program to avoid double-counting the same results.
- Fine-tuning accuracy changes are expressed in percentage points; diagnostic peak F1 is distinct from accuracy.
- The diagnostic agent account includes negative ablations and preserves the distinction between answer-conditioned teacher experiments and independent diagnostic performance.
- Optimization runtime gains retain the simulated-latency qualification. The convergence result belongs to the prior Combettes–Eckstein algorithm; this work contributes implementation and experiments.
- The Zuse work links to its Optimization Online preprint (posted April 2025; updated May 2026), the current public PDF, and AsyncProx. The current public PDF and updated local copy supersede the older draft timings: 881.144 to 728.214 seconds (17.4%) for two workers and 923.442 to 701.419 seconds (24.0%) for three. Repository availability does not establish journal acceptance.
- The controversy report supersedes the old 89% resume claim: the best reported prompting result is 87%, while text-plus-edit-count logistic regression reaches 93.1% on only 29 test examples. The WebNLG account credits Aryan with the from-scratch GPT-2 pipeline, first GCN method, preprocessing, and shared writing; final BLEU scores are team results and unsuccessful embedding experiments are retained.
- The thesis contains unfinished appendix placeholders and some literature-versus-experiment ambiguities. The portfolio uses results explicitly reported in the main experimental chapters and distinguishes those from cited leaderboard results; it does not claim prospective clinical validation.
- Abacus skills work describes the infrastructure framework enabling users to create and use their own skills with their own DeepAgent instances, as clarified by Aryan.
- Lymph content includes Aryan's subsequent supplied account of LangGraph/LangChain, OpenSearch/Qdrant, citation provenance, Langfuse/evaluation tooling, Cloud Tasks, and Cloud Run/Terraform. Durable operations and replayable SSE replace earlier blanket resumability/checkpoint claims. Adoption figures and institutional references are supplied by Aryan.
- Substack links and the M.Tech GPA were removed from website content at Aryan's request. The GPA remains in the downloadable resume. The old graph-to-text repository URL returned 404 and was removed; the project account remains.
The application resume now uses an original, modern single-column design with an oversized name, restrained teal/navy palette, sans-serif body type, measured emphasis, and white space. The master CV uses the same visual system, deliberate page composition, running headers, and page numbers. Both PDFs preserve selectable text and portable hyperlinks. The one-pager is selective; the CV carries the full account.
Theme references reviewed visually: RenderCV Classic, Modern-Deedy, and Jake's Resume. The implementation is original HTML/CSS rather than a copy of template source.
- Google resume overview: consistent formatting, impact, and relevant evidence in the top half.
- Amazon recruiter guidance: simple text-forward presentation and quantified accomplishments.
- Anthropic hiring guidance: foreground interesting research and open-source contributions.
The supplied research PDFs are linked directly for readers to inspect. GitHub Pages publishes the repository root from main at https://tan90cot0.github.io/. Pushes to main update the existing site. Both the one-page resume and detailed CV are available from the home page.