Repository lifecycle: Active → Maintained → Experimental/Incubating → Archived. Component repos stay separate only when independent versioning, reuse, access control, CI, or deployment boundaries justify it. See PORTFOLIO.md.

• agentic-vault — agent-agnostic knowledge infrastructure for Obsidian, including graph-aware retrieval, provenance, and multi-agent governance
• agentic-analytics — MCP-native analytical runtime with reproducible execution and evidence provenance
• Agent infrastructure (agent-*) — host-neutral, independently usable layers for running AI agents under real constraints: agent-relay (roles, handoffs, verification), agent-router (cost-aware model and executor routing), agent-control (lifecycle, budgets, policy, audit), and agent-toolchain (distribution across harnesses)
• litert-lm-plugin-cc — local/on-device LiteRT-LM integration for Claude Code
• DIMER model pipelines — model integrations for the DIMER platform spanning tabular, vision, detection, document AI, vision-language, generation, Earth observation, motion and matching, the life and physical sciences, time-series, audio, and language tasks, each pinning one immutable model revision, verifying weight checksums before load, and shipping machine-readable provenance with every result

• market-study-template — a scaffold for structured market research studies: a Quarto book rendering to PDF/DOCX, a TypeScript survey instrument, and pre-specified analysis contracts so the output tables exist before any data is collected
• citation-audit — audits a finished reference list against the bibliographic record, catching fabricated, misattributed, and misdirected citations — including identifiers that resolve cleanly to an entirely different paper. Keyless; the calls that change what a document claims stay with the human
• agentic-research — 24-skill PRISMA 2020 / GRADE systematic-review pipeline: a question → a defensible synthesis
• agentic-analytics — an agent-agnostic, MCP-native analytical runtime: reproducible sandboxed execution with source→execution→evidence provenance and deterministic validation, so an agent's analysis can be re-run and checked instead of taken on trust

• keynote-builder — a governed, human-gated pipeline that turns a brief into a stage-ready keynote (slide deck + speaker script)
• office-reports — agent skills for institutional document drafting: travel reports, minutes, and a weekly → terminal M&E reporting pipeline, evidence-only with gap flags
• presentation-builder — composable skills for finished presentations: Marp, native PPTX, HTML slides, talk scripts
• research-deck-builder — a research/training module → a presenter-ready, source-checked deck
• agentic-vault — an Obsidian vault template for working with AI agents (PARA + multi-agent governance)

• ai-readiness-assessment — an AI-readiness assessment chatbot for Philippine government agencies & NGOs
• rice-pest-detector — a teaching demo for counting rice black bug and white stemborer in light-trap photos with RT-DETR, RF-DETR, and D-FINE on CPU; research and training only, not field-validated

• litert-lm-plugin-cc — an unofficial Claude Code plugin for a local, on-device model served by LiteRT-LM: offline, no API tokens, nothing leaving the machine — and a metadata repair for the silent CPU fallback that costs a measured 4.3×
• model-dependabot — Dependabot-style alerts for AI model dependencies: weight files, locked packages, base models, and training data, checked against advisories and upstream retractions or licence changes. Read-only; it never touches the model
• DIMER model pipelines — model integrations for the DIMER platform, all built to one contract: a pinned immutable model revision, checksum-verified weights, validation before inference, machine-readable provenance on every result, and a model card recording intended use, out-of-scope use, and decision thresholds. Each repository publishes its own release status; most sit at candidate/source-complete pending recorded clean-runtime evidence
  • Tabular — Mitra, TabICL, TabPFN, TabDPT classifier/regressor pairs
  • Vision — Swin classification, detection, and segmentation; SAM 2 and SAM ViT promptable segmentation and CLIPSeg text-prompted segmentation; Depth Anything V2 relative and ZoeDepth metric depth; ViTPose human keypoints; X-CLIP zero-shot video classification; DINOv2 features; Mask2Former panoptic and Mask R-CNN instance segmentation; MODNet portrait matting; MediaPipe Face Landmarker face geometry (no identification); Swin2SR 2× and 4× super-resolution and NAFNet deblurring; ViT-MAE masked-autoencoder pretraining; and ViT / DeiT / CvT / MaxViT / EVA02 / ConvNeXt / ConvNeXt V2 / EfficientNet / ResNet-50 / MobileNetV4 classifiers
  • Detection — Grounding DINO, OWLv2, and OWL-ViT open-vocabulary detection from text prompts; CountGD open-world counting from text and visual exemplars; DETR, Conditional DETR, YOLOS, and RT-DETR set-prediction COCO detection; and YOLOX-S / YOLOX-X anchor-free detection with bounded SimOTA fine-tuning
  • Document AI and OCR — GOT-OCR 2.0 plain and formatted text recognition; SmolDocling page image → layout, reading order, and tables; LayoutLM and Pix2Struct document question answering; Table Transformer table detection and structure recognition; DePlot chart-to-table; DiT document-type classification; and Pix2Struct diagram QA, text-aware captioning, and UI widget captioning
  • Vision-language — SigLIP 2, Florence-2, SmolVLM, and BLIP captioning, visual question answering, and image-text matching; SigLIP v1 zero-shot classification and retrieval; and domain CLIPs BioCLIP 2 for species and PubMedCLIP for biomedical images (research only, not a clinical tool)
  • Generation — FLUX.1 [schnell] and PixArt-Σ text-to-image; Kandinsky 2.2 text-to-image, inpainting, and depth ControlNet; and Kandinsky 5.0 Video Lite text- and image-to-video
  • Earth observation and climate — Prithvi-EO flood and burn-scar segmentation, crop classification, and HLS feature extraction; CFNet bi-temporal change detection; and Aurora global weather forecasting
  • Motion and matching — RAFT optical flow and stereo disparity; LightGlue sparse and XoFTR detector-free image matching
  • Life and physical sciences — ESM-2 protein embeddings; Nucleotide Transformer genomics; Geneformer and scGPT single-cell transcriptomics; MoLFormer molecular properties; and MACE interatomic potentials for materials
  • Time series — Chronos-2 and TiRex / Toto / TimesFM forecasting; MOMENT embeddings, imputation, and anomaly scoring
  • Audio and multimodal — Whisper ASR, AST audio classification, Kokoro TTS, and Phi-4 multimodal
  • Language — LoRA/QLoRA fine-tuning contracts for compatible causal LMs, GLiNER NER, Qwen3 embedding and reranking, plus BERT / GPT-2 / T5 baselines
  • Text tasks — extractive question answering over SQuAD2, table question answering over WikiTableQuestions, abstractive summarization, zero-shot classification via NLI entailment, and English→Tagalog translation

• Agent infrastructure (agent-*) — four independently usable layers for running AI agents under real constraints: agent-relay hands work between agents through durable artifacts and a verifier gate; agent-router routes each subtask to the cheapest executor that can satisfy it, then escalates on failed verification; agent-control owns lifecycle, budgets, policy, and an append-only audit trail; agent-toolchain installs and validates all of it across harnesses