# Claude Sonnet 4 Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/claude-sonnet-4
> Markdown URL: https://aitinkerers.org/technologies/claude-sonnet-4.md
> Technology record last updated: 2026-02-24T17:02:38Z
> Generated: 2026-09-21T13:44:11Z

Anthropic's mid-tier model: balances state-of-the-art coding (72.7% on SWE-bench) with cost-efficiency for high-volume, agentic workflows.

Claude Sonnet 4, released May 22, 2025, is Anthropic's versatile, mid-tier model (successor to Sonnet 3.7). It delivers superior coding performance, achieving a state-of-the-art 72.7% on SWE-bench. The model features hybrid reasoning: it offers both near-instant responses and an 'extended thinking' mode for deeper analysis. Optimized for high-volume applications, Sonnet 4 provides an optimal balance of capability and practicality, with pricing set at $3/$15 per million tokens (input/output). Access is readily available via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI.

- Official technology site: https://www.anthropic.com/claude
- Public AI Tinkerers demos and talks: 18
- Result page: 1 of 1

## Recent Public Talks and Demos

### [Katalog — an AI app that makes post-surgery shoulder range of motion visible](https://tokyo.aitinkerers.org/talks/rsvp_BDzQthaV_gg)

Katalog is an AI app that supports rehabilitation after shoulder surgery by making range of motion visible. I broke my shoulder. Six months after surgery I found out my arm wasn't going up — I had thought it was. A mirror shows you where your hand reaches, not what your shoulder actually did. Prop a phone against something, stand 2 m back, raise your arm for 60 seconds. It reports the angle measured against your own body axis, plus how much your body compensated to get there. Pose estimation runs inside the browser; no video ever leaves the device. Live I'll show the working system end to end: - measuring (the screen says 135°, the record says 107° — the trunk was leaning) - mirror vs. measurement side by side, same instant, same hand height - the history, and the agent reading it and deciding one action As a stretch goal I'm rebuilding the attention layer in integer arithmetic (CIA) to run ViT on an FPGA and measure its power draw — if that lands in time, I'll show that version too.

- Event context: AI Tinkerers Tokyo — October 5 Meetup — 2026-10-05 — Tokyo
- Public talk page: https://tokyo.aitinkerers.org/talks/rsvp_BDzQthaV_gg

### [A Self-Improving AI OS for the Chronically Self-Employed](https://barcelona.aitinkerers.org/talks/rsvp_AA_yQ6WeHtc)

A personal AI operating system, built as a plain-text git repo, that connects to my entire freelance business in real time: live contracts, open invoices, job feed, inbox, calendar, and meeting notes. It uses that context to find work, draft proposals, track clients, and run a morning brief every day without me asking. This system is built upon the foundation of Nate Herk's AIS-OS combined with Andrej Karpathy's LLM-wiki guidelines. For the demo I'll show it live: open a terminal, pull the job feed from Contra, trigger a skill, watch the wiki update, then walk through the self-audit loop that finds gaps in the system and ships one improvement per week. The repo stays visible throughout. No slides. We may jump into live client / prospects on Contra.

- Event context: AI Tinkerers Barcelona - September Demo Night — 2026-09-17 — Barcelona
- Public talk page: https://barcelona.aitinkerers.org/talks/rsvp_AA_yQ6WeHtc

### [From 1,157 handwritten forms to one geological fact: a 25 m pile-driving limit](https://curitiba.aitinkerers.org/talks/rsvp_nt6LgCapvyQ)

A resumable extraction pipeline that turned 1,157 handwritten pile-driving field bulletins — scanned into 1,400+ messy PDFs — into a clean, georeferenced dataset, and then into engineering knowledge. The arc is the point: reports → data → knowledge. Live, I'll show the raw handwritten forms going in (smudged, non-standard, with voided sheets and unrelated pages mixed in), parallel subagents transcribing each to JSON, and the consolidated 1,157-record CSV driving an interactive report — histograms, box plots, an E×N heatmap, and a 3D scatter of 627 georeferenced piles. Then the payoff: those production-control reports became a technical and geological analysis instrument. Aggregated, the field data no single form could show revealed a consistent practical driving limit of ~25 m for the site (mean driven depth 24 m) — a hard geological fact that makes the next foundation estimate far more assertive. The dataset is a real completed job: 1,000+ pre-cast concrete piles driven over six months for a grain terminal in southern Brazil.

- Event context: AI Tinkerers Curitiba: Encontro de Agosto (no EBANX) — 2026-08-26 — Curitiba
- Public talk page: https://curitiba.aitinkerers.org/talks/rsvp_nt6LgCapvyQ

### [The Prompt That Worked in Isolation and Died in Production: Eval-Driven Prompt Engineering for a Real-Time Interview Copilot](https://seoul.aitinkerers.org/talks/rsvp_S4LRW_RGb3U)

InterviewMate (interviewmate.tech) is a real-time interview copilot I built and run in production: it listens to a live interview, transcribes it as it happens, and streams back answer suggestions grounded in the candidate's own experience — in under a second. The pipeline is Deepgram streaming STT into a FastAPI WebSocket Q&A pipeline, RAG retrieval over my own past answers, and a Claude streaming response. I'll run the working system live. Then I'll show the part that actually mattered: the eval harness behind the system prompt. It started as a bug — my copilot answered the viral "car wash" question correctly ("drive") while the same model, called with no prompt around it, said "walk," and I couldn't explain which prompt layer was responsible. So I turned the production prompt into a controlled experiment: 6 prompt conditions, 20 runs each, showing STAR-style structured reasoning beating context injection by 2.83x — and a follow-up where that same winning prompt collapsed to 0–30% once I dropped it into the real 60-line production prompt. Code, eval logs, and the pipeline architecture will all be on screen.

- Event context: AI Tinkerers Seoul: Inaugoral Demo Night — 2026-07-02 — Seoul
- Public talk page: https://seoul.aitinkerers.org/talks/rsvp_S4LRW_RGb3U

### [Selling to Scientists: Sales Intent Identification for Super Technical Buyers](https://seattle.aitinkerers.org/talks/rsvp_zsBLWVQ1iE4)

Sales intent tools watch behavior — they tell you an account is researching your category. They can't tell you what a researcher is working on, what they're stuck on, or what would make them reply. I sold to AI researchers for a decade by reading their papers and asking questions only someone who read them would ask. I built the system I wish I'd had the whole time: a seller defines their product, target accounts, and ideal buying signals in plain English; the engine mines papers, linked repos, and adjacent sources, then surfaces named opportunities with openers grounded in real pain quotes from the buyer's own writing. The novel part isn't the retrieval. It's that the seller's domain expertise becomes the system's input — captured as natural language signals, not keywords or trained models — and the LLM applies that expertise across a corpus the seller could never read on their own.

- Event context: GTM / Growth Track - Seattle — 2026-05-27 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_zsBLWVQ1iE4

### [JEN-R8 Discovery Engine: An AI That Tries to Prove Itself Wrong](https://manchester-nh.aitinkerers.org/talks/rsvp_1xgFaiUwGiA)

JEN-R8 is an AI research engine that tries to prove itself wrong. It synthesizes hypotheses across scientific domains, pre-registers the exact conditions under which it will declare itself defeated, runs the analysis on public data, and auto-generates a post-mortem when its own gates fail. Those post-mortems then feed back into the next round of hypothesis generation; the engine learns from its own failures. You'll see the full loop live: cross-domain bridge -> drafted hypothesis spec -> pre-registered falsification gates -> real-time execution on public data -> a meaningful failure -> the engine articulating, on stage, exactly why its own claim was wrong. Then I'll walk through the operator console. Current state (as of demo time): - 60 hypotheses across 15 scientific domains - 5 hypotheses confirmed through full gate passage - 6 hypotheses abandoned with documented post-mortems - 5 provisional patents filed in cancer biomarker discovery - 3 pre-print papers in preparation - 2 developing collaborations with university research institutes JEN-R8 was designed and built by agents running on the Wisdom Layer SDK (the subject of my AI Tinkerers demo last month, featured in the April global newsletter.)

- Event context: AI Tinkerers Manchester (Bedford), NH - May 2026 Meetup — 2026-05-20 — Manchester NH
- Public talk page: https://manchester-nh.aitinkerers.org/talks/rsvp_1xgFaiUwGiA

### [AI for people who still print their email: how we put a 25-tool agent in front of 50-year-old accountants without a single hallucinated delete](https://poland.aitinkerers.org/talks/rsvp_d3AEfot6ruQ)

Numonis is an accounting SaaS in Spain and Portugal where the primary users are 40–60-year-old SMB owners and accountants, and we shipped a pydantic-ai conversational agent with 25+ tools (invoicing, documents, banking, P&L, tickets) as the main interface. Live demo: I run a demo of the system's features all over one typed SSE stream. Between beats I flip to DevTools to show the raw frames so the wire protocol is visible end to end.

- Event context: AI Tinkerers Poland #3 - Meetup in Wrocław — 2026-05-06 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_d3AEfot6ruQ

### [AI meets our Analog world in Audio Digital Signal Processing](https://dublin.aitinkerers.org/talks/rsvp_fw5WukO9a4Q)

Audio-cleaner: An audio digital signal processing program with frequency filtering and echo cancellation. While AI tool quickly found opensource and textbook algorithms and pulled them into the architect's plug-in solution, it needed help making the solution cross platform and hallucinated bugs when the core problem was that endienness is not standardised in Microsoft Windows.

- Event context: AI Tinkerers Dublin: Demo Night with Fin — 2026-04-28 — Dublin
- Public talk page: https://dublin.aitinkerers.org/talks/rsvp_fw5WukO9a4Q

### [Harness: A Multi-Agent Claude Orchestrator That Builds and QA's Games Autonomously](https://singapore.aitinkerers.org/talks/rsvp_RwBQJVJwoLk)

Harness is a game-agnostic multi-agent orchestrator that uses the Claude Agent SDK to autonomously build and test games end-to-end. It reads a feature-list.json from any target game project, spawns a role-specific generator agent (gameplay / UI-art / audio) to implement each feature with full tool access (Read, Write, Edit, Bash, Glob, Grep + Godot MCP), then spawns a separate evaluator agent that launches the game, classifies each test step as DATA / VISUAL / BOTH, captures screenshots, and emits pass/fail verdicts. A watchdog loop detects stuck sessions and a pre-flight smoke test validates tool permissions before real work begins. Everything streams live to a single-process Fastify + React dashboard over WebSockets — you watch agents ship and break features in real time. For the demo I'll drive Harness against a live Godot project, show the full loop (pick → implement → evaluate → retry), and walk through how session records, screenshots, and eval reports are persisted back to the target repo.

- Event context: AI Tinkerers Singapore: The Agentic Future & Dev/Eng Workflows — 2026-04-21 — Singapore
- Public talk page: https://singapore.aitinkerers.org/talks/rsvp_RwBQJVJwoLk

### [How Personalized AI Outputs Turned Our Internal Tool Into a Product](https://chicago.aitinkerers.org/talks/rsvp_V18d7zsohTA)

Food Blog Studio is a SaaS that provides food bloggers with 10+ AI powered tools to speed up their day to day workflows. Demo context: Walkthrough of the full app, show off how personalization works and how it matches the users voice/tone/context, the transition from "internal tool" to "tool with users"

- Event context: AI Tinkerers Chicago: April Meetup ft OneTwoLoop — 2026-04-14 — Chicago
- Public talk page: https://chicago.aitinkerers.org/talks/rsvp_V18d7zsohTA

### [Lazy Marketing v3: From Meeting to Content Pipeline in Zero Click](https://seattle.aitinkerers.org/talks/rsvp_TtXOReDjy68)

Most of us have 10+ meetings a week full of insights worth sharing — but turning those into content never happens. Marketing Machine fixes that by wiring together Claude, Gemini, Slack, Google Sheets, and the LinkedIn API into a fully automated content pipeline. A Read.ai transcript hits a webhook. Claude extracts the best marketing hooks and expands them into on-brand LinkedIn posts using a 100+ line brand-voice prompt. You pick and refine posts entirely inside Slack with buttons and modals — no separate app needed. Gemini generates an accompanying image. Approved posts queue up in Google Sheets and auto-publish to LinkedIn on a schedule. In this demo, I'll walk through the full loop live — from raw meeting transcript to a published LinkedIn post — and dig into the prompt engineering, multi-model orchestration, and Slack-as-a-UI patterns that make it work. Built with Node.js, Express, and APIs anyone can sign up for. No GPUs, no fine-tuning, no ML infra required.

- Event context: AI Tinkerers Seattle: GTM Track — March — 2026-03-26 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_TtXOReDjy68

### [Cache-Optimized Agentic Fanout: Squeezing 65x More Intelligence Per Dollar From Claude](https://chicago.aitinkerers.org/talks/rsvp_0Ru5Zk7StVE)

We got a 65x reduction in per-rule input cost by designing the prompt architecture around Anthropic's cache economics — and the constraints that forced are more interesting than the savings. Live demo of a production system that evaluates 50+ natural-language compliance rules against 100+ page financial documents. Each rule needs full document context (~32k tokens), so naive execution is a non-starter on cost and rate limits. The fix: append-only message threads where nothing is ever mutated or removed, because any change to the prefix — including the tool schema — invalidates the cache. Tools that should not fire are rejected with a reply prompt, not removed from the definition. A warming request primes the cache before the real fanout begins. pg-boss orchestrates the map-reduce: Phase 1 parses documents in parallel via Reducto, Phase 2 fans out N independent agent invocations (one per rule, configurable concurrency per ECS task), Phase 3 reduces to review status. After the first cache write, subsequent rules pay ~500 new tokens instead of ~32k (!!!). The agent has four tools that persist directly to Postgres with no post-processing step: Think (structured reasoning), Calculate (mathjs-backed arithmetic verification), AnnotationWrite (bounding-box PDF evidence traceable to specific parsed document blocks), and RecommendationWrite (PASS/FAIL with confidence). Full OTel instrumentation with GenAI semantic conventions traces every tool call, token count, and cache hit/miss.

- Event context: AI Tinkerers Chicago: March Meetup ft. Programmers Inc. — 2026-03-17 — Chicago
- Public talk page: https://chicago.aitinkerers.org/talks/rsvp_0Ru5Zk7StVE

### [rm -rf salesforce — Claude Runs Deal Flow Now](https://seattle.aitinkerers.org/talks/rsvp_qa7m-04lrlg)

I built a git-backed knowledge base that replaces both my CRM and Google Docs for running an early-stage fund. Every entity — founders, companies, deals — lives as a folder of markdown and YAML. Claude agents are first-class participants: they process inbound emails, run entity resolution, create records, log notes, and commit directly to the repo. A GitHub Actions cron runs nightly, calling Claude Sonnet to synthesize each entity's raw files into a structured materialized summary. I'll demo the full live workflow: forward a pitch email → agent resolves the entity → creates the records → generates a materialized view → commits to git. No database. No SaaS. Just files, Claude, and git.

- Event context: AI Tinkerers Seattle: GTM Engineering Kickoff — Building AI for Growth — 2026-02-26 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_qa7m-04lrlg

### [Survey of Software: Mapping Python's Library Landscape with AI-Powered Research](https://seattle.aitinkerers.org/talks/rsvp_OA9n1DYQS8M)

A live research repository demonstrating systematic library discovery using Claude Code. Visit my station to explore 40+ completed categories (sorting algorithms, graph libraries, ML frameworks, crypto tools, etc.) organized in a Dewey Decimal-style taxonomy. I'll walk visitors through: The research output - Browse the Docusaurus site (https://research.modelcitizendeveloper.com/survey/) showing completed surveys The Four-Pass Survey methodology - How each category gets researched: Rapid Discovery → Comprehensive Analysis → Need-Driven Discovery → Strategic Selection Live demonstration - Run a new category research session with Claude Code if there's interest The metaprompt - Show the actual prompts that drive the research Visitors can: Request categories they need researched (128 still pending!) See how to adapt the methodology for their own domains Grab the metaprompt to run their own surveys Discuss the broader question: "What does development look like when AI can generate code instantly?" (Answer: Systematic library selection becomes the bottleneck) This is a working research project, messy and evolving. Perfect for science fair show-and-tell.

- Event context: AI Tinkerers Seattle: January Meetup — 2026-01-31 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_OA9n1DYQS8M

### [How to lose in fantasy football (A tutorial for programmers)](https://nyc.aitinkerers.org/talks/rsvp_wIsF1SMces8)

I'll be presenting a full-stack MCP application that enables Claude to autonomously manage ESPN Fantasy Football teams through natural conversation. This system demonstrates advanced MCP patterns including real-time tool execution monitoring, optimized context management (achieving 77% faster response times through static week context injection), and comprehensive API integration (36 specialized tools spanning roster management, trade proposals, waiver claims, and live player analytics). The live demo will showcase Claude analyzing rosters, making lineup decisions, and executing transactions against real ESPN leagues—illustrating the end-to-end workflow from user query to API write operations.

- Event context: NYC October Demo Day — 2025-10-02 — New York City
- Public talk page: https://nyc.aitinkerers.org/talks/rsvp_wIsF1SMces8

### [Building Working Code Live: Documentation-First AI Development](https://seattle.aitinkerers.org/talks/rsvp_ctw_B2iINIQ)

I'll demonstrate the spawn-experiments methodology - a systematic approach to AI-assisted code generation using documentation-first prompting with Claude models (Sonnet 4 and Opus) through Claude Code's Task tool. Live demo: building a functional CLI tool from scratch in real-time, showing how structured specifications generate better AI outputs than vague prompts. Based on spawn-experiments research comparing 4 AI development methodologies across multiple experiments, findings show methodology significantly impacts AI development patterns and outcomes, with trade-offs between development speed and code quality. You'll see the actual "code-like prompting" techniques that treat prompts as executable specifications, parallel code generation across multiple threads, and comprehensive quality control including syntax validation, automated test generation, and real-time error handling.

- Event context: AI Tinkerers Seattle September Meetup — September 30, 2025 — 2025-10-01 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_ctw_B2iINIQ

### [Intrinsic Dimension: Finding Optimal Dimensionality for Vector Embeddings](https://austin.aitinkerers.org/talks/rsvp_7Kk2MUDsi-0)

A Jupyter notebook with Python code that measures the intrinsic dimension (https://en.wikipedia.org/wiki/Intrinsic_dimension) of two large datasets of 768-dimensional semantic vector embeddings downloaded from Hugging Face. Interpretation of surprisingly low but complex results. Practical suggestions for how to use this method to choose the dimension of embedding vectors in various applications.

- Event context: Community AI Demos – Austin • July 10 2025 — 2025-07-10 — Austin
- Public talk page: https://austin.aitinkerers.org/talks/rsvp_7Kk2MUDsi-0

### [A Simple production Ready RAG pipeline](https://dublin.aitinkerers.org/talks/rsvp_ILfCX6RtjnQ)

A production ready implementation of RAG showing step by step how RAG with qdrant can be done and coming out of it knowing the fundamental basics of RAG

- Event context: AI Tinkerers - Dublin Event (June) — 2025-06-26 — Dublin
- Public talk page: https://dublin.aitinkerers.org/talks/rsvp_ILfCX6RtjnQ

## Related Technologies

- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [FastAPI](https://aitinkerers.org/technologies/fastapi) ([Markdown](https://aitinkerers.org/technologies/fastapi.md)) — 181 public demos
- [Claude Code](https://aitinkerers.org/technologies/claude-code) ([Markdown](https://aitinkerers.org/technologies/claude-code.md)) — 215 public demos
- [Amazon Bedrock](https://aitinkerers.org/technologies/amazon-bedrock) ([Markdown](https://aitinkerers.org/technologies/amazon-bedrock.md)) — 20 public demos
- [Claude Haiku](https://aitinkerers.org/technologies/claude-haiku) ([Markdown](https://aitinkerers.org/technologies/claude-haiku.md)) — 10 public demos
- [Claude Opus 4](https://aitinkerers.org/technologies/claude-opus-4) ([Markdown](https://aitinkerers.org/technologies/claude-opus-4.md)) — 4 public demos
- [GitHub Actions](https://aitinkerers.org/technologies/github-actions) ([Markdown](https://aitinkerers.org/technologies/github-actions.md)) — 20 public demos
- [MCP](https://aitinkerers.org/technologies/mcp) ([Markdown](https://aitinkerers.org/technologies/mcp.md)) — 129 public demos
- [Next](https://aitinkerers.org/technologies/next) ([Markdown](https://aitinkerers.org/technologies/next.md)) — 186 public demos
- [Qdrant](https://aitinkerers.org/technologies/qdrant) ([Markdown](https://aitinkerers.org/technologies/qdrant.md)) — 17 public demos
- [AGENTS](https://aitinkerers.org/technologies/agents) ([Markdown](https://aitinkerers.org/technologies/agents.md)) — 20 public demos
- [Anaconda](https://aitinkerers.org/technologies/anaconda) ([Markdown](https://aitinkerers.org/technologies/anaconda.md)) — 1 public demo
- [Anthropic](https://aitinkerers.org/technologies/anthropic) ([Markdown](https://aitinkerers.org/technologies/anthropic.md)) — 36 public demos
- [@anthropic-ai/claude-agent-sdk](https://aitinkerers.org/technologies/anthropic-ai-claude-agent-sdk) ([Markdown](https://aitinkerers.org/technologies/anthropic-ai-claude-agent-sdk.md)) — 1 public demo
- [Anthropic API](https://aitinkerers.org/technologies/anthropic-api) ([Markdown](https://aitinkerers.org/technologies/anthropic-api.md)) — 66 public demos
- [arXiv API](https://aitinkerers.org/technologies/arxiv-api) ([Markdown](https://aitinkerers.org/technologies/arxiv-api.md)) — 1 public demo
- [AWS Bedrock](https://aitinkerers.org/technologies/aws-bedrock) ([Markdown](https://aitinkerers.org/technologies/aws-bedrock.md)) — 9 public demos
- [Claude](https://aitinkerers.org/technologies/claude) ([Markdown](https://aitinkerers.org/technologies/claude.md)) — 173 public demos