dawson. Get in touch

LLM products and agent systems

Applied AI

I build LLM features people use every day: content pipelines that check their own output, tutors, analysts and hosted agents. The model is one part; the checks, data and operations around it are the work.

Email me about this workSee all work

What ties this work together

  • Output that gets checkedGenerated questions, reports and scores are verified against answers, rubrics or schemas before anyone sees them.
  • Agents that run every dayScheduled agent pipelines collect, analyze, write and publish without hand-holding, and their output is reviewed.
  • Any model that fitsClaude, GPT, Gemini, DeepSeek and local models behind one routing layer, chosen per task for quality and cost.

Top work

Ranked by how complete and substantial each one is, flagship first.

Both News screenshot

Flagship

Both News

Nonprofit news site: Claude scores each story for political lean and reporting balance, an influence tracker follows narratives, and every page ships in many languages.

15.8Karticles analysed

57outlets tracked

9languages

  • Claude returns a structured analysis per story: lean score, six balance ratings, key facts, missing perspectives
  • Headlines refresh every two hours from a maintained source list; results cached by content hash
  • Influence tracker follows narratives, profiles publishers and shows which voices each side amplifies
  • Many languages with per-language URLs, a browser extension API and first-party analytics
  • Next.js 16
  • Anthropic Claude API
  • Supabase
  • Vercel
  • Browser extension
  • Live
  • Built solo
GoTrails screenshot

GoTrails

Trail search in plain language, an AI briefing on whether to go, trail Q&A and a weekly weekend plan built from location, history and the forecast.

494Ktrails

171countries

23languages

  • Natural-language search over 494K trails, e.g. "Saturday morning with the kids"
  • AI go/no-go briefing per trail from the forecast, heat, shade and steepness
  • Ask-about-this-trail answers and a weekly personal weekend plan
  • Vue 3
  • MapLibre GL
  • Supabase
  • Kotlin + Jetpack Compose
  • Live
  • Built solo
Shotory screenshot

Shotory

One sentence to a finished video: Claude scripts, real AI video with lip-sync, consistent cast, automatic frame QA.

  • One sentence in, finished vertical video out: Claude writes the shot-by-shot script
  • Real AI video routed across Google Veo 3.1, Kling 3.0 and OmniHuman, with lip-synced dialogue
  • Same faces, sets and products in every shot via reference frames; a vision model grades and redraws weak frames
  • Vue 3
  • Supabase
  • Claude
  • Google Veo 3.1 on Vertex AI
  • Live
  • Built solo
MyClaw screenshot

MyClaw

Hosting platform that provisions a dedicated AI-agent server per user across four clouds, with DNS, SSL, a browser terminal and billing.

  • Provisioning on AWS Lightsail, Hetzner, DigitalOcean and Fly.io with cloud-init
  • Hono API that manages instances over SSH; chat, files, skills and channels from the dashboard
  • Extended from an open-source MIT project with a multi-provider layer
  • Hono
  • Drizzle ORM + PostgreSQL
  • Supabase
  • AWS Lightsail / Hetzner / DigitalOcean / Fly.io APIs
  • Open source
  • Built solo
WhyNext screenshot

WhyNext

LLM agents write daily market reasoning as schema-validated JSON; a deterministic renderer publishes it, and every call is graded after the close.

268daily briefs published

  • LLM agents write the daily report as JSON validated against a schema
  • A deterministic template renders it, so the model never writes HTML
  • Each scenario and trade idea is graded in the evening review
  • Python
  • Pydantic
  • Jinja2
  • Polygon.io market data
  • Live
  • Built solo
ReadLift AI screenshot

ReadLift AI

AI reading coach for grades 3–12: a Socratic tutor whose conversations become a skill profile aligned to state standards.

  • Socratic tutor Sage asks questions instead of handing out answers
  • Conversations become structured skill signals in a reading profile
  • Read-aloud with word-by-word highlighting, voice input, six languages
  • Next.js 16
  • Vercel AI SDK
  • Anthropic Claude
  • Supabase
  • Live
  • Built solo

More AI work

10 more, in ranked order.

  1. 7

    SmartSAT

    An SAT prep platform whose question bank is written, checked and illustrated by LLM pipelines, with an AI tutor on top.

    smartsat.us

    • TanStack Router & Query
    • Supabase
    • OpenRouter
    • Stripe
    Visit
  2. 8

    WriteLift AI

    Essays graded against the state rubric by an AI pipeline, with an evaluation suite that checks scoring quality.

    • Next.js 16
    • Supabase
    • OpenRouter
    • shadcn/ui
    Visit
  3. 9

    BizScout

    Staged agent pipeline (collect, score, write, maintain) that publishes a daily AI-opportunity report.

    • OpenClaw / Hermes agent runtime
    • OpenRouter
    • Python
    • Web search & source scraping
    Visit
  4. 10

    HouseFlipScore

    AI agent writes a per-parcel investment analysis on top of explainable scoring of county data.

    houseflipscore.com

    • Next.js 16
    • Prisma
    • PostgreSQL
    • Auth.js v5
    Visit
  5. 11

    Brand X-Ray

    Agent-produced brand audits: cross-platform data collection, benchmarks and a 100-point health score.

    brand-x-ray.com

    • OpenClaw agent workspace
    • LLM-driven research & analysis
    • HTML → PDF reports
    • Vercel
    Visit
  6. 12

    MyClaw Desktop

    Electron desktop client for a CLI agent runtime: multi-agent chat, channels, cron automation and skills.

    github.com/microblue/myclaw-desktop

    • Electron 40
    • OpenClaw
    • Zustand
    • i18next
    Code
  7. 13

    One-click Gemma

    One-click local LLM install (Rust + Tauri) that wires Gemma into an agent runtime through an OpenAI-compatible API.

    github.com/microblue/one-click-gemma

    • Rust
    • Tauri 2
    • Tokio + reqwest
    • Bash & PowerShell installers
    Code
  8. 14

    OkText

    Remote MCP server with OAuth that lets ChatGPT and Claude send and read texts from the owner's phone, approved on the device.

    api.oktext.app

    • Go
    • SQLite
    • Kotlin
    • HTTPS long-poll
    Visit
  9. 15

    VWork

    Experiment in a 'virtual company' of AI agents defined in YAML, with a task board and daily standups.

    github.com/microblue/vwork

    • Python 3.12
    • Click
    • Rich
    • PyYAML
    Code
  10. 16

    ralphmon

    Terminal dashboard for monitoring and controlling autonomous Claude Code loops across projects.

    github.com/microblue/ralphmon

    • Python 3.11+
    • Textual
    • psutil
    • tmux
    Code

Other areas

The rest of the work, by area.

Fintech & quant5 projectsAI education5 projectsFull-stack products11 projectsAll workEvery project in one list