Models are easy. Systems are the job.

A monocular SLAM pipeline built from scratch — ORB extraction, essential-matrix initialization, PnP tracking, sparse bundle adjustment, and g2o pose-graph loop closure — wrapped in a Flask dashboard for running and inspecting real dataset runs. Validated against ground truth on TUM (0.119m ATE) and KITTI (36.2m ATE over 200 frames of real driving, reported honestly as expected monocular drift, not hidden).

A fine-tuned classifier paired with LLM reasoning and a built-in disagreement check. YOLOv8-cls trained on ~54,000 PlantVillage images identifies crop and disease across 38 classes; below 60% confidence it asks for a better photo rather than guessing. The label goes to Gemini for a structured treatment plan, and the original image goes separately to Gemini vision for an independent second opinion that flags symptom-prediction mismatches.

An n8n workflow that turns an uploaded invoice into a validated database record with no manual entry. A webhook accepts PDF or image and branches by type — text extraction or base64 for Claude Vision — converging on a Claude call returning structured JSON. Validates required fields, checks PostgreSQL for duplicates, returns 201 / 409 / 422 accordingly. Claude request bodies are built as real expression objects, so quotes and newlines in invoice text can't break the payload.

Pose estimation feeding pure geometry. YOLO11-pose extracts 17 body keypoints per frame; trigonometry, no second model, derives spine angle, knee flex, arm extension, shoulder tilt, hip tilt, and X-factor, the shoulder-to-hip separation that predicts power transfer. Video mode overlays a live skeleton and charts how the angles evolve across the swing.

Claude's vision model reads an actual product photo, not a filename or category tag, and writes usable marketing copy in under ten seconds. Four styles crossed with four tones, a 50–350 word length control, and .txt export.

A research agent architected as a LangGraph state machine: recall prior findings from ChromaDB, plan sub-queries, search and scrape the web, self-evaluate evidence sufficiency with a retry loop and hard iteration cap, then synthesize a cited Markdown report and persist a summary for future recall. Memory-assisted queries ran ~33% faster with ~40% fewer searches than cold-start ones in benchmark runs.

An autonomous-vehicle perception console running a Cityscapes-finetuned SegFormer-B0 on driving scenes, with pixel-level overlays across 19 classes and a connected-components instance approximation clearly labeled as such rather than passed off as true instance segmentation. Styled as a safety console — KPI strip, safety-critical object count, class groups that put people, vehicles, and traffic signals ahead of background classes. The rebuild fixed a real torch/torchvision version mismatch and a false-positive misclassification where tree canopies were being read as "person."