STRIDE

LaTeX. Tailored. ATS. Shipped.

• Scrapes live job postings from Greenhouse, Lever, and Internshala — plus deep-link buttons to LinkedIn, Indeed, Glassdoor, Wellfound, and Naukri.
• Tailors your LaTeX resume to any JD: Gemini rewrites bullets and skills to weave in keywords while preserving your formatting.
• Emits a PDF compiled server-side by Tectonic, auto-shrunk to one page, named {FullName}_{Role}.pdf.

Stack — React + Vite + Tailwind · FastAPI · Tectonic (LaTeX → PDF) · Google Gemini (free tier) · BeautifulSoup (Internshala scraper).

Live at https://getstrideai.vercel.app (frontend) · https://stride-backend-zbza.onrender.com/api/health (backend).

---

Project layout

Catalyst2.0/
├── backend/     FastAPI service: tailor pipeline + job search
└── frontend/    React + Vite app (cream + black premium theme)

---

What's in the app

Route	Purpose
/	Landing — hero video, intro, pipeline, how-to-use
/app	Tailor flow: paste .tex + JD → PDF download
/search	Search live boards + deep-link buttons to LinkedIn/Indeed/Glassdoor/Wellfound/Naukri

Default-resume presets — both pages offer three built-in starting points (Off-Campus, On-Campus, Priority / high-priority off-campus). Click one to load it into the editor; edit it and hit ★ SAVE to persist your own version of that preset (stored per-preset in the browser under localStorage["stride:default:<id>"], never sent to the backend). A • marks a preset you've customized.

---

Prerequisites (local dev)

1. Python 3.11+ (tested on 3.14)
2. Node 18+
3. Tectonic — see Windows install below; Linux/Mac use your package manager
4. A free Gemini API key — https://aistudio.google.com/apikey

Installing Tectonic on Windows

winget doesn't have it. Use the prebuilt release:

$dest = "$env:USERPROFILE\tectonic"
New-Item -ItemType Directory -Force -Path $dest | Out-Null
Invoke-WebRequest `
  -Uri "https://github.com/tectonic-typesetting/tectonic/releases/download/tectonic%400.16.9/tectonic-0.16.9-x86_64-pc-windows-msvc.zip" `
  -OutFile "$dest\tectonic.zip"
Expand-Archive -Path "$dest\tectonic.zip" -DestinationPath $dest -Force
Remove-Item "$dest\tectonic.zip"
& "$dest\tectonic.exe" --version

Then create a fonts.conf next to the binary (Tectonic's Windows build needs fontconfig pointed at a config file or it crashes with Cannot load default config file):

@'
<?xml version="1.0"?>
<!DOCTYPE fontconfig SYSTEM "urn:fontconfig:fonts.dtd">
<fontconfig>
  <dir>C:/Windows/Fonts</dir>
  <cachedir>~/.cache/fontconfig</cachedir>
  <config><rescan><int>30</int></rescan></config>
</fontconfig>
'@ | Set-Content -Encoding UTF8 "$env:USERPROFILE\tectonic\fonts.conf"

The backend auto-detects the fonts.conf next to the Tectonic binary and sets FONTCONFIG_FILE + HOME in the compile subprocess.

Linux / Mac

Tectonic is available via package managers (apt, brew, cargo). Fontconfig is already configured on those systems, so the auto-detect simply finds nothing and Tectonic uses the system default.

---

Backend setup

cd backend
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
Copy-Item .env.example .env
# Edit backend\.env:
#   GEMINI_API_KEY=...                                (required)
#   TECTONIC_BIN=C:\Users\<you>\tectonic\tectonic.exe (Windows; or just "tectonic" on PATH)
uvicorn app.main:app --reload --reload-include "*.env"

--reload-include "*.env" makes uvicorn pick up .env edits without a manual restart.

Health check: http://localhost:8000/api/health

---

Frontend setup

cd frontend
npm install
npm run dev

Open http://localhost:5173. In dev, /api is proxied to localhost:8000.

---

Usage

The landing page's "How to use" section spells out the flow. Short version:

1. Save your default resume — on /app, paste your .tex once and click ★ SAVE. Stored in browser localStorage; future visits reload it with one click.
2. Find roles — /search filters live Greenhouse / Lever / Internshala boards by role, location, and internship status. For LinkedIn / Indeed / Glassdoor / Wellfound / Naukri, use the "More job sites" deep-link buttons (they open each platform's pre-filled search in a new tab — no scraping needed).
3. Tailor a match — click Tailor on any search result → jumps to /app with the JD pre-filled and your default resume already loaded.
4. Generate the PDF — click Generate. Gemini rewrites bullets and skills with JD keywords, Tectonic compiles, auto-shrunk to one page. ~15–30 seconds end-to-end. Downloads as {FullName}_{Role}.pdf.
5. Apply on the company site — STRIDE never submits the application form. Click Open Posting on any search result to land on the company's actual page, upload the tailored PDF, hit submit yourself.

---

API endpoints

POST /api/tailor

{ "latex_source": "...", "job_description": "..." }

Pipeline:

1. Pre-process the user's .tex to strip constructs that crash Tectonic on Windows (see caveats below).
2. Gemini rewrites bullets and skills using the JD keywords, preserving structure and bullet count. Returns a metadata JSON line + the new .tex.
3. Finalize the LLM output: same sanitizers, strip LLM-added \textbf{} from bullet bodies, repair/re-bullet list structure (orphan \resumeItem, flattened Achievements, broken \resumeSubHeadingListStart nesting), and verify the achievement-point count survived the rewrite.
4. Tectonic compiles at natural 11pt first. If the page overflows — two pages, or content clipped past the bottom (Overfull \vbox) — it escalates a one-page shrink: 10pt → 9pt → 8pt with progressively tighter line spacing, recompiling until it fits. Resumes that already fit keep their natural size (no sparse, over-shrunk look).
5. If a compile fails, a repair pass sends the broken .tex + the Tectonic error back to Gemini and retries once.
6. The PDF streams back with Content-Disposition: attachment; filename="FullName_Role.pdf".

The /api/tailor handler is a sync def so its blocking work (Gemini SDK + Tectonic subprocess) runs in FastAPI's threadpool — otherwise a single worker would block the event loop, starve /api/health, and get the instance restarted mid-request (a 502).

POST /api/search-jobs

{
  "role": "python developer",
  "location": "remote",
  "internship_only": true,
  "source": "",
  "top_n": 50
}

Fetches every free source concurrently (Greenhouse + Lever boards + Internshala scraper). Filters with progressive relaxation — strict role+location pass first, then drop-location pass, then drop-role pass — so the panel always returns ≥10 results when any exist. Tech queries get the CSE filter: non-tech Internshala jobs (Marketing/Sales/Real Estate) are stripped, and the Internshala URL routes to the reliable Software Development category instead of role-specific URLs that fall back to general WFH listings.

Boards results are cached in-memory for 10 minutes per server instance — first search per window pays the ~15–30s network cost, the rest run in ~3s.

GET /api/health

{ "status": "ok", "model": "gemini-2.5-flash" }

---

Caveats — what the pre-processor auto-fixes

Tectonic is stricter than most LaTeX engines and has a known Windows bug around FontAwesome. The pre-processor in backend/app/services/preprocess.py auto-fixes the most common gotchas:

Issue	Fix
\usepackage{fontawesome5} + \fa* commands crash Tectonic 0.16.x on Windows with heap corruption	Stripped entirely; ~ separators in \begin{center} blocks become `$
\input{glyphtounicode} + \pdfgentounicode=1 (pdfTeX-only, Tectonic uses XeTeX)	Stripped
LLM writes \end{resumeItemListStart} instead of \resumeItemListEnd	Regex post-fix: any \end{...Start} → \...End
LLM adds \textbf{} around bullet keywords despite the no-emphasis rule	Stripped from inside \resumeItem{...} bodies and skill value lists; structural bolds (job titles, category labels) preserved
Achievements render with no bullets — LLM dropped the list wrapper or flattened the section into a skills-style block	Re-bulleted: orphan \resumeItems get wrapped in \resumeItemListStart … End, and flattened achievement blocks are exploded back into \resumeItem bullets (title-scoped, so Technical Skills stays intentionally bulletless)
\resumeItemListStart nested directly inside \resumeSubHeadingListStart (no \resumeSubheading between) crashes Tectonic with "Something's wrong--perhaps a missing \item", zero pages out	The redundant outer wrapper is collapsed to a plain \resumeItemListStart … End item list (valid, still bulleted)
Resume overflows one page after tailoring	Conditional, escalating one-page shrink driven by page count + Overfull \vbox detection: compile natural 11pt, then 10pt → 9pt → 8pt with progressively tighter line spacing until it fits. Resumes that already fit are left at natural size. Applied only on the final compile, never before the LLM sees the source.
Filename gets "Internship", "Part Time", "Full Time" suffixes	Trailing employment-type tokens stripped from the role segment of the filename

---

Configuration reference (backend/.env)

GEMINI_API_KEY=...                                   # required
GEMINI_MODEL=gemini-2.5-flash                        # primary
GEMINI_MODEL_FALLBACK=gemini-2.5-flash-lite          # auto-fallback on 503/UNAVAILABLE; empty to disable
TECTONIC_BIN=C:\Users\you\tectonic\tectonic.exe      # or "tectonic" if on PATH (Linux/Mac default)
CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173,https://getstrideai.vercel.app
HOST=0.0.0.0
PORT=8000

---

Swapping the LLM provider

All Gemini calls funnel through backend/app/services/llm.py. To switch to OpenAI, Anthropic, or Groq: replace the _call_gemini() body and the system-prompt plumbing — the response parser (metadata JSON + ```latex code block) is provider-agnostic. Update requirements.txt and `.env` to match.

---

Deploying

Step-by-step deploy guide → DEPLOY.md (Vercel for the frontend, Render for the backend Docker image, with git push auto-deploys).

Quick summary:

• backend/Dockerfile builds a production image with Python + Tectonic (Linux static binary, no fontconfig setup needed unlike Windows).
• render.yaml provisions the backend on Render in one click.
• frontend/vercel.json configures Vercel as a Vite SPA.
• The two are linked via VITE_API_BASE (frontend → backend URL) and CORS_ORIGINS (backend allows frontend domain).

---

Production notes

• Lock CORS_ORIGINS in the backend's environment to your real frontend domain.
• Bundle the Tectonic binary into your container image and run as non-root (the included Dockerfile does this).
• Frontend builds to static files with npm run build — deploys cleanly to Vercel / Netlify / Cloudflare Pages.
• The free Gemini tier rate-limits aggressively (~15 RPM for gemini-2.5-flash); the LLM client retries on 503 / 429 / 500 up to 3× with exponential backoff.
• Render free tier sleeps after 15 min of inactivity → first request after sleep takes ~30–50s. Upgrade to Starter ($7/mo) if you want always-on.