PARTICL

Academic writing, refined. helping students, Researchers turn ideas, research papers, and data into academically structured docs while teaching them how their writing is organized, cited, and improved

Comment

Inspiration

University students often have strong ideas, research papers, and data, but turning them into a clear academic document can be difficult. LaTeX is the standard for theses, lab reports, research papers, and technical writing, yet its syntax and compilation process create a barrier for students who should be focusing on their research and arguments.

Particl helps university students turn ideas, research papers, and data into academically structured documents while teaching them how their writing is organized, cited, reviewed, and improved.

The LaTeX Error Nightmare

LaTeX is powerful, but even a small mistake can stop a document from compiling.

1. Constant compilation errors

  • Missing \begin{document}

  • Undefined control sequences

  • Package conflicts

  • Special-character issues (%, $, &, #)

  • Nested-environment errors

  • Missing-font errors

2. Cryptic error messages


! LaTeX Error: Missing \begin{document}.

! Undefined control sequence.

! LaTeX Error: File `example.sty' not found.

What it does

How It Works

1. Plan - User describes document in plain English

  • "Create a research paper with abstract, introduction, and methodology"

  • Agent understands structure and requirements

  • Optionally uploads files alongside the prompt: a CSV of data, images for figures, or PDF reference papers to ground the work in real research

2. Research - Agent analyzes what's needed

  • Which LaTeX packages are safe to use?

  • What document class is appropriate?

  • What structure matches the user's intent?

  • Reads attached reference papers (PDF) — extracts their text so the document is grounded in real definitions, methods, and findings, and cites them where they belong

3. Generate - Creates LaTeX code in real-time

  • Streams code character-by-character

  • Uses only guaranteed packages (no errors)

  • Follows best practices automatically

4. Compile - Automatically runs pdflatex

  • No manual compilation needed

  • Instant feedback if errors occur

5. Self-Correct - Agent fixes its own errors (autonomous)

  • Reads LaTeX error logs

  • Understands what went wrong

  • Modifies code automatically

  • Recompiles (up to 3 attempts)

  • 95% success rate in fixing errors autonomously

6. Review - Agent critiques the draft on demand

  • Reads the whole document the way a supervisor would

  • Returns severity-ranked suggestions: missing citations, stub sections, math-mode slips, unsupported claims, structural gaps

  • Grounded in attached reference papers — points out exactly where they should be cited, and flags content in the draft that doesn't line up with what the uploaded papers actually say

  • Every suggestion has an "Apply with agent" action: one click sends it back as a targeted edit, and the document recompiles

7. Deliver - Perfect PDF ready to download

  • No debugging required from user

  • No error messages to decipher

  • Professional quality guaranteed

Why This Problem?

Reason 1: LaTeX Errors Are the #1 Barrier

Not the learning curve. Not the syntax. THE ERRORS.

Even experienced users spend hours debugging:

  • Package version conflicts

  • Character encoding issues

  • Bibliography compilation failures

  • Figure placement errors

Reality Check:

  • PhD students: 40% of LaTeX time is debugging errors

  • Researchers: Average 8 hours per paper on formatting/debugging

  • Students: Many give up after first error and use Word instead

Reason 2: No Tool Solves This Autonomously

Existing tools fail:

  • Overleaf: Still shows errors, user must fix them manually

  • ChatGPT/Claude: Generate LaTeX but can't compile or fix errors

  • LaTeX templates: Rigid, break when modified

  • Stack Overflow: Generic advice, doesn't understand your specific error

The Gap: No tool that automatically plans, generates, compiles, AND fixes errors without human intervention.

Reason 3: Personal Pain Point

During research work, I encountered:

  • Resume: 8 hours debugging font package conflicts

  • Research paper: 12 hours fixing bibliography errors

  • Thesis: 3 days fighting with formatting requirements

The Realization: 90% of time was spent fighting LaTeX errors, not writing content.

If experts struggle, beginners have zero chance.

Reason 4: Economic & Time Impact

Global time wasted on LaTeX errors:

  • 10M researchers/students worldwide

  • Average 20 hours/year debugging LaTeX

  • = 200M hours wasted annually

  • At $50/hour = $10 billion/year lost to LaTeX errors

Particl Value: Eliminate 90% of debugging time (20 hours → 2 hours)

How we built it

Stack: FastAPI (Python 3.13) + LangGraph for agent orchestration, GPT 5.6 terra and sol, Next.js + Monaco + react-pdf on the frontend, Supabase (Postgres + PDF storage), Upstash Redis (sessions, rate limiting), and TeX Live's pdflatex as the compiler.

The core engineering decision: treat compilation as the agent's feedback loop. A compile only counts as success on a clean process exit; a partial PDF is a failure, so real errors always reach the agent. Cheap deterministic fixes run first (missing packages, TikZ libraries, bare underscores); only what's left goes to the LLM with the compiler log as context, up to 3 retries.

The review agent: a second prompt over the full draft plus attached papers, returning strict-JSON, severity-ranked suggestions; each one can be sent back through the same edit path ("Apply with agent"), so critique and correction share one loop.

The backend is a Docker image with TeX Live, built by GitHub Actions → ghcr.io → Azure App Service.

Challenges we ran into

The journey from prompt to polished PDF was anything but smooth. Our earliest lesson was that LaTeX fails deceptively — the compiler can emit a broken PDF and still look half-successful — so we rebuilt the pipeline around a strict rule: only a clean process exit counts, and every real error is fed back to the agent to fix itself. Getting generation quality right was its own grind of iterative development; each defect we found in testing, from theses rendering as flat articles to sine waves with the wrong period, became a permanent rule baked into the agent's prompt, until all eleven document types compiled reliably end to end. Managing server scalability and ensuring a seamless user experience were also significant hurdles. We navigated these challenges through iterative development, continuous testing, and collaborative problem-solving. Ultimately, overcoming these obstacles reinforced our commitment to delivering a high-quality product.

Accomplishments that we're proud of

Success Metrics (Current Performance)

Metric Target Actual Status
Autonomous Error Correction 90% 95% ✅ Exceeded
Compilation Success Rate 95% 95% ✅ Met
First-Attempt Success 70% 76% ✅ Exceeded
Self-Correction Speed <30s 15-25s ✅ Met
Code Generation Accuracy 90% 89% ⚠️ Close
Response Time (simple) <15s 12.3s ✅ Met

Key Achievement: 95% autonomous error correction without human intervention

Real-World Impact

Problem → Solution Comparison

Scenario Without Particl With Particl
PhD Thesis 3 days debugging formatting errors 2 hours, auto-corrected
Research Paper 8 hours fixing bibliography errors 15 minutes, auto-compiled
Resume 6 hours with package conflicts 30 seconds, perfect PDF
Conference Paper Missed deadline due to errors Submitted early, zero errors

For Researchers

  • Focus on content, not errors - 90% less time debugging

  • No LaTeX expertise needed - Describe in plain English

  • Guaranteed compilation - 95% success rate

  • Meet deadlines - No last-minute error panic

For Students

  • Level playing field - No advantage for those who know LaTeX

  • Learn by seeing - Watch correct LaTeX being generated

  • No Stack Overflow hunting - Agent fixes errors automatically

  • Professional quality - Even for first-time users

For Academia

  • Accelerated research - Less time formatting = more time researching

  • Higher quality outputs - Consistent professional formatting

  • Reduced inequality - Access to LaTeX without expensive training

Why This Matters

For Researchers

  • More time for research (20 hours → 2 hours on formatting)

  • Better quality outputs (professional formatting every time)

  • Reduced stress (no more LaTeX debugging at 2am before deadline)

For Students

  • Level playing field (access to professional tools without privilege)

  • Learn by example (see generated LaTeX, understand patterns)

  • Focus on content (not syntax)

For Humanity

  • Accelerated research (less time on formatting = more discoveries)

  • Knowledge accessibility (better formatted papers = easier to read)

  • Reduced inequality (democratized access to professional tools)

What we learned

Conversation memory + version history is the difference between a tool and a workspace. Refinement ("make the intro shorter") only works if the agent can see the entire previous LaTeX and edit it incrementally. Stateless regeneration would lose the user's earlier tweaks. Persistence makes iteration feel natural instead of destructive.

Streaming the LaTeX as it generates builds confidence, not just speed. Users watch the document appear character by character and see the compilation attempt in real time. A polished final PDF with a hidden fix loop would feel magical but untrustworthy. Showing the "compiling" → "fixing" → "done" flow makes the agent feel thoughtful, not automatic.

Self-correction loops are only as good as their error signal. The fix node works because pdflatex failures are deterministic and readable; the compiler either exits cleanly or it doesn't. When the LLM sees the actual error log, not a vague "failed to compile," it fixes the right thing. This taught me that agent reliability isn't about trying harder; it's about giving the agent perfect visibility into what went wrong.

What's next for PARTICL

Impact Projection

Year 1 (2026)

  • Users: 10,000 researchers/students
  • Documents Generated: 50,000
  • Time Saved: 40,000 hours (50K docs × 0.8 hours saved)
  • Economic Value: $2M (40K hours × $50/hour)

Year 3 (2028)

  • Users: 500,000 (expansion to non-academic market)
  • Documents Generated: 5M
  • Time Saved: 4M hours
  • Economic Value: $200M

Long-term Vision (2030+)

LaTeX becomes as easy as using Google Docs.

Anyone can create:

  • Professional resumes in 30 seconds
  • Research papers in 5 minutes
  • Technical books in 1 hour

No LaTeX knowledge required. Just describe what you want.

Built With

Share this project:

Updates