2025PRINTIMIZE — GRADUATION PROJECT

Print parameters, solved

A hybrid AI system for FDM 3D printing. Claude reads renders of your model with forced tool-use, so its geometry analysis comes back as a schema, never as prose to be parsed. A physics-constrained gradient-boosting model — trained on profilometer readings from 108 real printed samples — predicts surface roughness. A multi-objective Bayesian optimizer searches the parameter space against it, and PrusaSlicer validates the winners into print-ready G-code.

2025 — 2026 · ROLE: SOLE DEVELOPER · LIVE IN PRODUCTION · FASTAPI · REACT · THREE.JS · SCIKIT-OPTIMIZE

FIG 01THE PIPELINE
STLGEOMETRYMULTI-ANGLERENDERSCLAUDE VISIONFORCED TOOL-USEML MODELR² = 0.99BAYESIAN OPTgp_minimize · qNEHVI{ "overhang": 0.42, "layer_defect": true, "recommend": [...] }SCHEMA-BOUND OUTPUT — NO FREE-TEXT PARSINGPRUSASLICERCLI VALIDATIONG-CODEPRINT-READYPRINT FEEDBACK → CONFORMAL RECALIBRATION
STL → RENDERS → CLAUDE VISION → ML MODEL → BAYESIAN OPTIMIZER → SLICER → G-CODESWIPE →
TAB 01RESULTS

0.99

R² — ROUGHNESS MODEL

0.51 µm

RMSE

108

PROFILOMETER SAMPLES

30+

PRINTER PROFILES

TAB 02AI ENGINEERING DECISIONS

FORCED TOOL-USE FOR STRUCTURED OUTPUT

  • Claude analyses multi-angle renders of the STL and must answer through a tool schema
  • Overhangs, layer-defect risk and a parameter range come back typed, not as free text
  • No prose parsing, so a wording change upstream cannot break the pipeline

MULTI-OBJECTIVE BAYESIAN OPTIMIZATION

  • scikit-optimize gp_minimize plus a qNEHVI engine, Pareto post-filtering
  • Surface quality traded against print time — the operator picks a point on the front
  • Cheap ML screening first, expensive slicer validation only on the finalists

PHYSICS-CONSTRAINED ML

  • Gradient boosting with monotonic constraints — layer height cannot improve roughness
  • Trained on real lab measurements, not simulated data
  • A geometry fingerprint retrieves similar past prints for a warm start

A MEASURABLE FEEDBACK LOOP

  • Every recommendation can be answered with "did it actually print well?"
  • That ground truth feeds the failure classifier and conformal recalibration
  • Every billable model call is logged with tokens and latency, so cost can be recomputed when prices move
TAB 03STACK

AI / VISION

Claude (Anthropic) · forced tool-use · multi-angle STL renders

ML / OPTIMIZATION

scikit-learn · scikit-optimize · qNEHVI · conformal intervals

BACKEND

FastAPI · Python 3.11 · PostgreSQL · SQLAlchemy 2 · Alembic

FRONTEND

React 18 · TypeScript · Three.js · React Three Fiber · Zustand

SLICER: PRUSASLICER CLI · DEPLOY: DOCKER ON AWS EC2 · GRADUATION PROJECT, FSMVÜ