Skip to content

jdpinetta.ink — portfolio / 2026

Jonathan “Nagui” Pinetta

Product architect, technical writer, and AI systems builder.

I design documentation systems, localization workflows, AI-powered tools, and product interfaces that turn complex operations into usable systems.

View work Contact
  • AI systems
  • documentation
  • localization
  • developer tools

01 / Selected work

Featured projects

Nine systems built end to end — from problem framing and architecture to working product. Execution platforms, localization engines, evaluation tooling, and operator consoles.

Execution System

Dark Factory

AI product engineering platform — from design interview to spec, wireframe, pipeline run, and published artifact.

Problem
AI-assisted product work fragments across chat, docs, and tickets — leaving specs untracked, runs unreviewable, and decisions undiffable.
Built
A full-stack platform covering the design-to-delivery loop: interview sessions, spec synthesis, wireframes, pipeline runs, quality scoring, and artifact publishing — all anchored to run IDs.
Why it matters
Treats every product decision as a traceable, replayable run — not a chat log. Specs diff. Runs compare. Artifacts ship.
  • Multi-agent
  • Spec-first
  • Full-stack
Product page
Migration Control

Migrai

Local-first platform for analyzing, migrating, and monitoring AI-powered systems across model vendors — with verified behavioral evidence.

Problem
AI features quietly couple codebases to one vendor — and "what would switching take?" gets answered by guesswork instead of evidence.
Built
A local console with a deterministic coupling scanner (every finding cited to file and line), capability graph, compatibility checks, migration planner, reversible apply/rollback runtime, and an event-sourced execution monitor — 19 milestones, each closed with specs, ADRs, and verification reports.
Why it matters
Treats vendor migration as an engineering operation: scan → assess → plan → apply → validate, with rollback by construction and 450 tests plus 15 e2e scenarios behind it.
  • Vendor-neutral
  • Evidence-first
  • Local-first
Product page
Agent Governance

AEO Kit + AEO Console

Schema, CLI, and console tooling for governing agent-facing interfaces.

Problem
Agent tools and policies drift when teams define them through scattered prompts and config files.
Built
A TypeScript monorepo with schema, core build/audit/policy engine, CLI, and OpenAI/MCP-oriented emitters, paired with a local SvelteKit/Tauri console using SQLite and Drizzle.
Why it matters
Treats agent tools as managed interfaces that can be validated, tested, and operated locally.
  • Schema
  • CLI
  • Local-first
Product page
Platform Prototype

NexusAI Platform

High-fidelity AI operations console prototype with local-first data and orchestration UI.

Problem
AI platform concepts need credible IA and operator workflows before full backend investment.
Built
A SvelteKit prototype with dashboard, logs, models, jobs, chat, workbench, orchestrator canvas, sql.js client-side SQLite, optional Supabase sync, and test coverage.
Why it matters
Demonstrates platform thinking, operator UX, and a mock-to-cloud migration path without pretending to be a production backend.
  • Prototype
  • Operator UX
  • sql.js
Product page
Localization Engine

Omniglot

Localization systems for structured translation workflows, file ingestion, QA, TM, and glossary support.

Problem
Localization workflows break down around messy formats, context loss, inconsistent QA, and spreadsheet-based handoffs.
Built
Omniglot Wrapper for file-heavy import/export and validation workflows, plus Omniglot Next Gen as a spec-driven platform for projects, keys, documents, TM/glossary, QA, and guarded AI assistance.
Why it matters
Connects localization expertise with AI-assisted product engineering and operational workflow design.
  • Localization
  • Spec-driven
  • ETL
Product page
Local MLOps

TuneKit

Local fine-tuning platform — training runs, adapter registry, eval suites, and efficiency analytics in one operator UI.

Problem
Local fine-tuning is scattered across scripts, terminal logs, and ad-hoc notebooks — making runs hard to track, compare, or hand off.
Built
A full-stack MLOps UI with starter kits, dataset recipes, training run history, live metrics, a versioned adapter registry, eval suites, and a chat lab for model validation.
Why it matters
Brings production MLOps structure to local LLM experimentation — every run is reproducible, every adapter is versioned, every result is inspectable.
  • Fine-tuning
  • Adapters
  • Local-first
Product page
Signal Intelligence

Noise Distiller

Feed aggregation and intelligence platform — correlated event clustering, AI briefings, and synthesis pipeline for signal-heavy research.

Problem
Research and signal monitoring is fragmented across dozens of feeds, with no layer to correlate events, surface patterns, or generate structured briefings.
Built
A full-stack intelligence platform with multi-source feed ingestion, correlated cluster analysis, AI-generated briefings, a synthesis pipeline for artifact generation, and advisor and monitor workflows.
Why it matters
Turns a firehose of sources into structured intelligence — correlated clusters, briefings with source lineage, and synthesis artifacts that are replayable and comparable.
  • Feed aggregation
  • Intelligence
  • Synthesis
Product page
LLM Evaluation

Benchy

LLM code benchmark platform — real production corpora, multi-task evaluation, and per-model analysis with recall, hallucination, and decay metrics.

Problem
LLM benchmarks rely on synthetic datasets that don't reflect how models perform on real, dense production code — making model selection for engineering tasks largely guesswork.
Built
A benchmark runner and analysis platform using real production code corpora to evaluate models on recall-verbatim, compose, simplify, docgen, and repair tasks — with leaderboards, per-function drill-down, hallucination tracking, and decay curve analysis.
Why it matters
Grounds model evaluation in real-world code tasks. The analysis layer surfaces not just accuracy but hallucination rate, recall vs position, and per-function failure modes.
  • Benchmarking
  • Code corpora
  • Multi-model
Product page
Developer Tool

Night Loop

Spec-first project OS for agentic development — task lifecycle, work sessions, operational insight, and delivery metrics in a synthwave TUI.

Problem
Agentic development workflows fragment across terminals, docs, and chat — there's no unified surface for tracking tasks, sessions, specs, and risk signals in one place.
Built
A spec-first project OS with a synthwave TUI — tasks, work sessions, specs, context, insight, delivery metrics, review intelligence, and operational risk scoring — all local, filesystem-backed.
Why it matters
Treats the agentic dev loop as infrastructure: every session is logged, every task is traceable, and the TUI gives real-time operational health without leaving the terminal.
  • TUI
  • Spec-first
  • Local-first
Product page

Recurring patterns

Architecture decisions that repeat across these systems.

  • Spec-First Development

    Specs and plans drive work: Dark Factory runs, AEO interface artifacts, and Omniglot Next Gen’s authoritative specs tree.

    Dark Factory · Omniglot · AEO Kit + AEO Console

  • Model-Agnostic Architecture

    Dry runs, local LLMs, and swappable provider options where implemented — without hard-binding the whole stack to one vendor.

    Dark Factory · NexusAI Platform · TuneKit · Migrai

  • Local-First + Cloud Hybrid

    Workstations and SQLite-first paths, with optional cloud sync or hosting where the repo actually wires it.

    TuneKit · NexusAI Platform · AEO Kit + AEO Console

  • Human-in-the-Loop Validation

    Review-oriented flows: Omniglot QA in context, segment review, and SME-style checks.

    Omniglot

  • Deterministic Workflows

    Dry mode, CLI checks, and artifact-oriented pipelines so outcomes are inspectable instead of one-off chat.

    Dark Factory · AEO Kit + AEO Console · Migrai

  • UI as Operational Surface

    Consoles and prototypes built as operator surfaces — dashboards, jobs, tuning, settings — not slide-only narratives.

    NexusAI Platform · AEO Kit + AEO Console · TuneKit

02 / Capabilities

What I build with

  • AI product design

    Designing AI features around real workflows — agents, copilots, and pipelines that earn their place in the product.

  • Technical documentation systems

    Docs as infrastructure: information architecture, content pipelines, and knowledge bases that stay accurate as the product moves.

  • UX writing & microcopy

    Interface language that carries the system model — states, errors, empty screens, and the words between clicks.

  • Localization architecture

    Translation workflows, glossary systems, and batch pipelines that scale content across locales without losing control.

  • SvelteKit product prototyping

    From concept to working product surface fast — typed, componentized, and ready to put in front of real users.

  • Workflow automation

    Bots, schedulers, and integrations that remove the manual loop between tools people already use.

  • Agent & RAG system design

    Retrieval, grounding, and agent orchestration designed for traceability — answers you can audit, not just generate.

03 / Philosophy

How I work

  1. Systems before screens.

  2. Documentation is product infrastructure.

  3. AI should reduce operational drag, not add abstraction.

  4. Interfaces should expose truth, not decorative complexity.

04 / About

Behind the systems

I work across product, documentation, localization, and AI tooling. My focus is turning messy workflows into structured systems: knowledge bases, translation pipelines, automation tools, agent interfaces, and developer-facing products.

  • Product architecture
  • Technical writing
  • UX writing
  • Localization
  • AI tooling
  • Automation

05 / Contact

Let’s build a system

Available for product, AI tooling, documentation systems, and localization architecture work.