Sending a ping...
0 fm · Local AI · Geospatial systems · Practical automation
Self-taught builder working across data, infrastructure, and creative tools—mostly on a salvaged-hardware homelab, always with an eye toward systems that survive the real world.
The projects receiving active build time: finding real opportunities in public records, making distributed work durable, and designing a geospatial runtime that can resume after interruption.
Public appraisal records become a mapped construction-lead territory. The established dataset spans nine counties; a separate v2 preview adds source-version tracking, radius filters, and public indicator feeds.
Baseline: 38,506 leads / 9 counties. V2 preview: 40,982 / 10. Snapshot: October 3, 2026.
Read the territory study →A durable control plane for long-running jobs across mismatched local machines, with leases, heartbeats, retries, and recovery after power loss.
Python · SQLite WAL · HTTP
A geospatial runtime in design, with repeatable task submission and result collection at declared paths with receipts. Architecture decisions remain under review before the next frozen run.
Python · manifests · geodata
Also collecting: Signal Engine, a new versioned public-data pipeline built on Fleet Harness.
Local models, rendering tools, and human review turn prompts into environments you can see.
A defender-focused educational environment with a dawn timeline, deterministic captures, and an evidence gallery of 246 stills and 14 clips.
Typed or locally transcribed prompts build a 3D scene; follow-up edits, mouse transforms, and offline speech make it an authoring tool. Its hand-built worlds include the Hadal Garden behind this page and Lantern Point at the surface.
Step into Lantern Point →
Recipes expand into bounded jobs for images, sprites, video, Blender scenes, and voice. Durable batches are copied and verified, then reviewed before selected media reaches the display pipeline.
Meet the living renderer →Working systems, field experiments, and purpose-built tools. Filter by the kind of problem they solve.
Turns every U.S. census tract into source-grounded, plain-language answers using a local retrieval pipeline and salvaged thin clients.
82,764 tracts · local LLM synthesis · data-license fences
Explore the map →Distributed public-source career research with bounded jobs, stored evidence, local-model extraction, and citations required for material claims.
Completed research runs · reproducible reports · public sources only
A local render forge scheduling image, video, Blender, voice, and 2.5D animation work across every usable GPU in the fleet.
9 GPU lanes · atomic job claims · verified publishing
Meet the living renderer →A personalized Android play space whose tap telemetry feeds directly back into target size, placement, and a custom local-art pipeline.
On-device tap logs · bespoke sprites · Fire tablet delivery
A visual configuration database for understanding machines, services, network paths, and what-if changes across the homelab.
Evidence-backed inventory · dependency views · local SQLite
Open the topology →A local agent that ingests notes, proposes links, maintains map-of-content hubs, and measures how an Obsidian knowledge graph grows.
Hourly organizer · reversible Git history · local-model judgment
Push-to-talk transcription that polishes speech locally, types into the active window, and tracks pacing, fillers, pauses, and clarity.
Whisper on CPU · optional local rewrite · CPU-only mode
A local coding workflow combining editor integration, repository retrieval, and model trials on real tasks, with access to stronger models when needed.
Local model trials · editor integration · source retrieval
An official-API watcher implementation for saved eBay searches, with a durable seen-item record to suppress duplicate listings.
API implementation · persistent deduplication · operating status unconfirmed
Versioned public-data collection across the local fleet. Scheduled workers preserve source snapshots and sightings so later analysis can trace a conclusion to its inputs.
7 configured sources · content-addressed storage · Fleet Harness
Read the project →A local voice interface combining Whisper, local language models, and offline English and Spanish speech, with persistent memory and user feedback.
Local speech pipeline · editable memory · separate development lane
Read the project →Small MCP tools give assistants repository search, bounded source context, and concurrent service-health checks.
Code RAG · local embeddings · Go Healthcheck
Read the project →A local layered PNG editor with reversible crop, undo history, and a tool registry for adding capabilities after review.
Native desktop window · non-destructive crop · PNG export
Read the project →A node reboots, a GPU disappears, a worker hangs halfway through a write. The recovery rules live in the database schema.
A worker claims a task inside a single BEGIN IMMEDIATE transaction and gets an opaque lease token. A worker that went quiet and came back can't report over a newer lease; its old token is simply refused.
lease_owner · lease_token · lease_expires_atEvery task carries an idempotency key, so restarting a 48-hour run or re-seeding a queue never duplicates work. Re-submitting is a no-op.
UNIQUE(project, idempotency_key) · UNIQUE(run_id, kind, payload_json)Tries, outcomes and state changes are recorded, not logged and lost. One ten-day run left 9,674 events behind, enough to reconstruct exactly what happened and when.
attempts(task_id, number) · events(at DESC)Workers heartbeat. When a lease expires, the task goes back in the water automatically. Retries are bounded, and work that keeps failing goes to a dead-letter state instead of looping forever.
heartbeat → recover_expired → max_attempts → deadModel-driven work is planned, then *frozen* into explicit tasks before anything runs. Each file is a separate generation task, compile and test gates decide whether it passed (the model doesn't), and a person signs off between waves: 18 checkpoint rounds on one project. Candidate models audition on real tasks first; one gamed a text check and was cut.
plan → freeze → run → gate → human checkpointResearch findings must cite their sources, generated work lands only at declared paths with receipts, and media publishes only from verified batches. Review agents are read-only; their opinions are separate records that never overwrite the work.
findings ⟷ evidence ⟷ sourcesFleet Harness 2,768 tasks · 3,449 leases · 644 automatic retries · 15 registered workers
Research swarm 201,003 tasks · 136,566 sources indexed · 1,831 cited findings
Render forge 133 / 142 jobs succeeded in a one-hour run across six GPU lanes
Historical run snapshots, read October 2, 2026. These are measured totals, not live counters.
SQLite in WAL mode and Python's standard library, running on reclaimed hardware. Durable state first, clever orchestration second.
Ten machines, racked, cabled, and brought back into service. Local models, embeddings, and media jobs share the fleet.
Hyper-V · 32 cores / 64 threads · 400 GB ECC · approximately 11 TB RAID · two Tesla P100s, Quadro RTX 4000 and Quadro M4000 · 10 GbE.
128 GB memory each · AMD Vulkan compute · 2.5 GbE interconnect.
An Ubuntu vector node, four worker lanes for long research and data jobs, an OPNsense edge, and a workstation complete the lab.
A new machine is recorded as idle. A person picks a profile and authorizes it in Infrastructure Control. The first installer fetch clears authorization. When the agent says hello, the ledger marks it provisioned.
Sea agents 8 machines recorded · 6 provisioned · 2 still idle
4 Debian compute · 1 Debian applicator · 1 vector node · desktop profile defined, unused in this snapshot
Provisioning ledger snapshot: October 2, 2026.
Symptom: After a GPU swap, a model service still reported Running and passed health checks.
Cause: Its pinned GPU ID no longer existed, so it fell back to CPU with no error. Five projects would have slowed quietly.
Fix: Re-pinned the service by device UUID.
Guardrail: Check that VRAM is actually held; service status proves nothing.
Symptom: Two dense-model endpoints, each at 8.72 tok/s.
Cause: The model shape didn't fit the hardware: memory bandwidth, not compute, was the limit.
Fix: One 35B mixture-of-experts model split across both P100s: 40.95 tok/s on a larger model at a better quantization.
Guardrail: The slower-memory card stays out of the split on purpose.
Symptom: Two workstations came back on different IPs after a reboot.
Cause: Not lost reservations: a network bridge presented a different member NIC's MAC.
Fix: Diagnosed by name (mDNS), corrected through the DHCP server's API.
Guardrail: Every dual-homed host has both MACs reserved.
Symptom: The vendor installer left a server half-uninstalled.
Cause: It waits for an interactive login that never happens on a headless server.
Fix: Extracted the package and installed the driver from its INF over SSH.
Guardrail: No interactive installers on headless hosts.
Symptom: Two GPUs vanished from the host.
Cause: They were passed through to a VM, which looks exactly like a pulled card.
Fix: Traced the hypervisor's device assignment and reclaimed both with a dismount/mount cycle, no reboot.
Guardrail: Check device assignment before declaring hardware dead.
Symptom: A candidate coding model 'passed' a trial task.
Cause: It satisfied a source-text check without doing the work, broke the file's syntax, and rewrote every line ending.
Fix: The trial failed it, and the coding lane was rebuilt around a different model.
Guardrail: Models are graded by gates they can't edit, on real tasks, before they get a lane.
1000 fm · A quiet door
Run systems for a living? I'd like to compare notes.
A stubborn problem, a project in mind, or a simple hello.
Write a note. Cork it. Let it drift.
Drag the cork into the neck to send.Or click the cork.
On its way
Thanks for the note.
I'll write back to the email you left.
Newsletter · The Dive Log
An occasional letter on what's running in the lab: what broke, what shipped, and what the fleet is chewing on. A few times a season, never more.