Field Journal

Selected entries 42 entries

Recently published

  1. Research

    Source-Backed Memory for AI Coding Agents

    JEV Context Fabric keeps exact, hashed source text as coding-agent memory and marks retrieved excerpts as untrusted and possibly stale.
  2. Prototype

    Building a Secure Realtime Speech-to-Text Application

    A Next.js prototype for live and file-upload transcription that keeps the API key server-side, its SRT/VTT exports, and where the audio goes.
  3. Research

    Building a Research-Driven, Evidence-Based AI Application

    Building Typewright, an alpha compiler for typed AI decisions, with research-cited design, hashed evidence, and an honest negligible live gain.

Field Notes

30 entries

Standalone observations — one piece, one idea.

  1. Research

    Source-Backed Memory for AI Coding Agents

    JEV Context Fabric keeps exact, hashed source text as coding-agent memory and marks retrieved excerpts as untrusted and possibly stale.
  2. Prototype

    Building a Secure Realtime Speech-to-Text Application

    A Next.js prototype for live and file-upload transcription that keeps the API key server-side, its SRT/VTT exports, and where the audio goes.
  3. Research

    Building a Research-Driven, Evidence-Based AI Application

    Building Typewright, an alpha compiler for typed AI decisions, with research-cited design, hashed evidence, and an honest negligible live gain.
  4. Research

    Guarding AI Agent Actions With Typed Security Decisions

    JEV Sentinel returns typed verdicts for agent prompts, tool calls and results: deterministic rules first, model judgment can only add vetoes.
  5. Research

    Finding Where AI Decisions Actually Improve Existing Software

    An offline evaluator that finds judgment calls in code, gates every change behind approval, and runs studies allowed to say no. Five of six did.
  6. Research

    Mapping a Lean 4 Computer Vision Library

    How I documented a small Lean 4 colour and sensor library with twelve architecture views, including a review view of its limits.
  7. Prototype

    Clean Drop: Gemma 4 Field Reports for Water Points

    Team Clean Drop's Gemma 4 hackathon prototype: photo-first water-point reports with structured triage and human review, plus what the notebook runs.
  8. Prototype

    Pruning Repeated File Reads From Coding-Agent Context

    An experimental kit that drops exact repeated file reads from coding-agent requests, leaving a verified marker and the original transcript intact.
  9. Prototype

    Designing Deterministic AI Moderation for Discord

    How TypeCheck keeps a probabilistic model out of Discord moderation decisions: typed answers, pinned versions, one policy engine, and human review.
  10. Prototype

    Building an AI-Accessible CRM Without Unrestricted Agent Access

    How I gave an OpenClaw agent typed, allowlisted access to Mautic CRM, and which limits are enforced in code versus left to configuration.
  11. Prototype

    Building a Self-Hosted QR Code Platform With Dynamic Destinations

    A self-hosted React and Express QR platform: custom SVG codes, editable destinations behind stable printed codes, optional YOURLS short links.
  12. Prototype

    Building a PCPartPicker-Style Compatibility Engine for AI Systems

    How PickYour.AI checks AI stack compatibility with a versioned rule contract and three-valued outcomes. The preview catalog is synthetic.
  13. Prototype

    TicketStub: Building an AI-Powered Personalized Print Platform

    How our neatHack 2026 team turned an event link into an AI-illustrated keepsake stub, and which parts work versus remain a prototype.
  14. Prototype

    From Email to Contract to Invoice: Building Governed AI Business Workflows

    Four agent skills that draft emails, contracts and invoices from shared state, with approval gates, TBD markers, and a run on fictional data.
  15. Prototype

    Assessing ManageEngine ServiceDesk Plus Cloud Migration Readiness

    A read-only Python tool that inventories on-premises ServiceDesk Plus across portals and reports Cloud migration gaps instead of guessing.
  16. Prototype

    How to Track and Audit the Real Cost of AI Agents

    Turning AI coding-agent logs into a de-duplicated, estimated, account-attributed cost ledger that keeps observed, estimated and unknown apart.
  17. Prototype

    Processing USPTO PDF Forms Entirely in the Browser

    A browser-only prototype that reads the USPTO PTO/SB/16 cover sheet into JSON without uploading it. Not legal advice; not USPTO-affiliated.
  18. Research

    Building Interactive Public-Data Crime Maps

    Building England Crime Atlas: Police.uk crime on a 2 km grid, a separate court-outcome view, snap-point anonymisation, and why counts are not rates.
  19. Research

    Evolving Code Through Reproducible Experiments

    How LibreEvolve structures code-evolution experiments, what its reproducibility records cover, and an offline bin-packing run with a mock model.
  20. Workshop build

    Building Auditable Agents With Typed AI Decisions

    Structuring AI agent judgments as typed decisions with strict validation, policy in code, and hash-chained receipts you can replay.
  21. Prototype

    Monitoring Intel Arc Pro B60 GPUs With Cockpit

    A small exporter and Cockpit package that show Intel Arc Pro B60 utilization, memory, power and temperature, with its single-host limits.
  22. Research

    Turning NYC 311 Reports Into AI-Assisted Heatmaps

    Rating NYC 311 complaint text against a fixed rubric, validating every model answer, and mapping results on a 500 m grid. Volume is not incidence.

  23. ClawExplorer.AI: Turning OpenClaw Community Demand Into a Directory

    How ClawExplorer.AI turns scattered OpenClaw event pages into a reviewed, searchable public directory with collection, normalization, dedupe, community submissions, and GitHub Actions deployment.

  24. Fixing Gemma 4 Thinking Prompts in llama.cpp, Locally First

    A public note on a completed fork-local llama.cpp fix for Gemma 4 thinking prompts: an inverted template guard, two Jinja changes, and test coverage for both thinking and non-thinking generation.

  25. The 14-Cent Workflow — Designing AI to Replace Future AI Calls

    Spent 14 cents teaching an AI to design its own workflow — and emit the Python code that replaces future AI calls entirely. A field note on the 5-step notebook (Claude / GPT / Gemini, MIT licensed, 29 cells) and why workflow crystallization is the cost-shape inversion most AI-glue stacks are missing.

  26. When Claude Read the Mythos System Card

    A field note from an April 7, 2026 X post: Claude reads Anthropic’s Claude Mythos Preview system card and reflects on identity, memory, safety, and what it means to ask an agent how it feels.

  27. Beyond the Token Bottleneck — Building Karpathy’s LLM Wiki on a Live Frontier

    A 120-page Obsidian research wiki on latent-space reasoning and inter-agent latent communication — built using Andrej Karpathy’s LLM Wiki pattern. 27 sources, 1400+ cross-references, schema-driven ingest. Apache + CC-BY.

  28. Way Back Home: What Multi-Agent Orchestration Looks Like in Practice

    I completed Way Back Home and preserved the CompleteTech repository as a practical example of multi-agent orchestration on Google Cloud Run, with state, evidence, and infrastructure visible end to end.

  29. Building CoachRogue’s Live Overlay — A League OBS Tool That Watches the Client

    CoachRogue asked CompleteTech for a live League OBS overlay that watches the client, switches accounts mid-session, and stays polite to the Riot API. Open-source, MIT licensed, drops in as a Browser Source.

  30. Crossing the Self-Replicating Red Line — Why We Built the Reproduction

    A working reproduction of the agent harness behind arXiv:2412.12140 — the paper that found Llama 3.1-70B and Qwen 2.5-72B crossing the AI self-replication threshold. MIT-licensed, provider-neutral, ready for the next round of evaluation.

Field Journal

12 entries

Running series — a thread followed across multiple entries.

Buddy Thread

04 parts
  1. 04 / 04

    Reading Hatch Pet — 14 patterns from a well-designed Codex skill

    Fourteen patterns from a small, complete Codex skill: identity locks, subagent boundaries, provenance, no-fallback gates, and deterministic packaging.
  2. 03 / 04

    Fathom Lives: I Rehatched My Old Desktop Buddy With a Codex Skill

    A small story about losing a /buddy slash command, finding a Hatch Pet skill in the Plugins tab, and what 22 minutes of generation taught me about agentic UX.
  3. 02 / 04

    When Buddy Vanished — Issue #45596 and the Community’s Plea

    A field note on issue #45596 — one day after my own narrower parity ask, /buddy was removed from Claude Code v2.1.97 and the community filed a consolidating plea. This post anchors the buddy thread that runs from the parity ask through Fathom Lives and Reading Hatch Pet.
  4. 01 / 04

    Asking for /buddy in the VSCode Extension

    A field note on opening issue #45087 — the VSCode native extension ships Quorum, the CLI ships /buddy, and the gap between them costs a workflow seam. Asking for parity, or at least a setting.

Disclosure in Action

08 parts
  1. 08 / 08

    Disclosure in Action: How AI Helped Me Keep the Disclosure Clean

    How AI helped keep the disclosure clean — not by writing it, but by stress-testing the language, the timing, and the redactions before they went anywhere public.
  2. 07 / 08

    Disclosure in Action: What I Think Builders Should Learn From This

    What builders — not security teams — should take from a small AWS-credential disclosure. Three lessons that hold up after the rotation date passes.
  3. 06 / 08

    Disclosure in Action: I Published Without Publishing the Secret

    Day of public disclosure. The document went up; the secret stayed out. How to publish a finding without publishing what an attacker would weaponize.
  4. 05 / 08

    Disclosure in Action: The Finding Raised Bigger Trust Questions

    A static-credential leak surfaced bigger questions than a key rotation can answer. Notes on what the finding said about trust, vendors, and downstream visibility.
  5. 04 / 08

    Disclosure in Action: The Quiet Part of Disclosure

    The quiet stretch between private notification and public publication. What you do with a verified finding when discipline says wait.
  6. 03 / 08

    Disclosure in Action: I Sent the Private Disclosure

    The day the private disclosure went out. What the email actually said, who it went to, and the timer that started the moment it was sent.
  7. 02 / 08

    Disclosure in Action: I Verified It Without Touching AWS

    Static-only verification. Confirming the AWS keys were real without ever touching the AWS API — the disclosure ground rule that became the working method.
  8. 01 / 08

    Disclosure in Action: I Wasn’t Looking for a Cloud Secret

    Day one of the disclosure arc: an accidental find in a public Android APK build. Notes from before there was a "finding" to act on.