Practical tools built with AI — assessment chatbots, agent skills, and lightweight ML systems — spec-driven and human-gated where it matters.
A ready-to-use scaffold for structured market research studies — a Quarto book that renders to PDF and DOCX, a TypeScript survey instrument, and pre-specified analysis contracts so the output tables exist before any data is collected.
View repositoryAudits 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 by default; the calls that change what a document claims stay with the human.
View repositoryAI-agent pipeline for literature and systematic reviews — 23 composable skills aligned with PRISMA 2020, Cochrane, and GRADE. Verifies every citation against the real bibliographic record, with human gates where LLMs are weakest.
View repositoryAn agent-agnostic, MCP-native analytical runtime — reproducible sandboxed execution with source-to-execution-to-evidence provenance and deterministic validation, so an agent's analysis can be re-run and checked instead of taken on trust.
View repositoryA governed, human-gated pipeline that turns a keynote brief into a stage-ready deck and speaker script end to end — capture through produce, pausing at the decisions only a human should make.
View repositoryAgent skills for institutional document drafting — official travel reports, minutes, and a weekly-to-terminal M&E reporting pipeline. Evidence-only, flagging gaps instead of inventing content.
View repositoryA suite of composable, agent-ready skills that turn source material into finished presentations — Marp research decks, native PowerPoint, HTML slides, spoken talk scripts, and producer rundowns. Each follows the open Agent Skills standard and chains into a source-to-deck pipeline.
View repositoryA skill that turns research modules into presenter-ready PowerPoint decks — source-checked citations, embedded speaker scripts, and a clean vector-shape design. Also repairs existing decks.
View repositoryObsidian vault template for working with AI agents — PARA + Zettelkasten organization, a multi-agent governance layer, 36 reusable agent skills, and hands-off automation with git snapshots and session logs.
View repositorySelf-service chatbot for assessing AI readiness of Philippine government agencies and NGOs — a guided, conversational evaluation that surfaces maturity, use cases, and a downloadable report.
View repositoryAn unofficial Claude Code plugin for calling a local, on-device model served by LiteRT-LM — offline, no API tokens, nothing leaving the machine. It also repairs the missing metadata key that silently drops published models onto CPU, a measured 4.3× penalty with no run-time override.
View repositoryA full-lifecycle, single-GPU MLOps platform that runs locally — data, train, register, serve, monitor, retrain — with one race-free GPU tenant, gated promotion, and a closed drift-to-retrain policy loop.
View repositoryModel integrations for the DIMER platform, all built to one contract: a pinned immutable model revision, weights verified by checksum before they load, dataset validation that runs ahead of inference, machine-readable provenance attached to every result, and a model card recording intended use, out-of-scope use, and decision thresholds. Coverage spans tabular, vision, detection, document AI, vision-language, time series, audio, and language — the last including question answering, summarization, and English–Tagalog translation. Each repository states its own release status; most sit at candidate / source-complete pending recorded clean-runtime evidence.
Tabular — Mitra, TabICL, TabPFN, TabDPT Vision — Swin, SAM 2, CLIPSeg, DINOv2, ViTPose Detection — Grounding DINO, OWLv2, RT-DETR, YOLOX Document AI — OCR, document QA, tables, charts Vision-language — SigLIP 2, Florence-2, SmolVLM, BLIP Time series — Chronos-2, TiRex, Toto, MOMENT Audio — Whisper, AST, Kokoro Language — LoRA/QLoRA, GLiNER, Qwen3 Text tasks — QA, summarization, zero-shot, translationFour independently usable layers for running AI agents under real constraints. Relay hands work between agents through durable artifacts and a verifier gate. Router sends each subtask to the cheapest executor that can satisfy it, then escalates on failed verification. Control owns lifecycle, budgets, policy, and an append-only audit trail. Toolchain installs and validates the set across harnesses.
agent-relay agent-router agent-control agent-toolchainA private, offline-capable PWA for signing PDFs — real PKCS#12 digital signatures with your signature image, in-app Digital ID creation, and counter-signing, all fully client-side so nothing ever leaves your device.
View repository