Practical tools built with AI — assessment chatbots, agent skills, and lightweight ML systems — spec-driven and human-gated where it matters.

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, in progress)
• aiaas-market-study — reproducible AI-as-a-Service market/feasibility research workflow (private)
• inventory-intelligence — applied inventory intelligence system (private)
• litert-lm-plugin-cc — local/on-device LiteRT-LM integration for Claude Code
• mlops-grid — MLOps infrastructure work (private)
• Mitra tabular pipelines — DIMER pipelines fine-tuning the Mitra tabular foundation model: classifier and regressor, each shipping a CPU dataset validator and a GPU fine-tuner

• 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 — 23-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

• 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×
• mlops-lite — a full-lifecycle, single-GPU MLOps platform (data → train → serve → monitor → retrain)
• Mitra tabular pipelines — DIMER pipelines that fine-tune Mitra (AutoGluon, Apache-2.0), a tabular foundation model pretrained on 45M synthetic datasets and no real data: classifier and regressor, each shipping a CPU dataset validator and a GPU fine-tuner
• TabICLv2 tabular pipelines (incubating) — DIMER pipelines for TabICLv2 (Inria Soda, BSD-3-Clause), an in-context learner that reads labelled rows as context: classifier and regressor
• notebooks — ready-to-run Colab speech & ML experiments (ASR, diarization, TTS)

• 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 (in progress) installs and validates all of it across harnesses
• pdf-signer-pwa — private, offline PKCS#12 PDF signing that never leaves your device