Hunter Karman

A multi-agent simulation harness that places language-model agents inside an access-controlled organizational environment and logs every resource request, communication, policy verdict, and unauthorized data flow.

To prevent label contamination, access decisions and leak detection are adjudicated by a deterministic policy engine rather than an LLM judge. Each run separates ground-truth state (events.jsonl, labels.jsonl) from the partial view available to external monitors (observable.jsonl), supporting both insider-threat detection benchmarks and agent-compliance evaluations across model families.

A local evaluation harness for comparing how audio-capable multimodal language models describe the same musical excerpt. Each run hashes the uploaded audio (WAV, MP3, FLAC, AAC, or OGG), encodes it as base64 input_audio, and dispatches identical system instructions, prompts, and sampling parameters across selected models with provider fallbacks disabled.

Responses, latency, token usage, and normalized musical descriptors are persisted in SQLite via Drizzle ORM and exported as JSON or CSV for blind side-by-side comparison.

A browser-based point-cloud renderer built on Three.js and custom GLSL shaders that reconstructs interior and street scenes from PLY spatial captures or procedural worker fixtures.

Rather than smoothing away capture defects, the shader pipeline exposes controls for scan-ring periodicity, depth quantization bands, missing-data wedges, layer misalignment, and temporal decay asseen regions thin out over time. Binary PLY parsing runs off the main thread in a Web Worker and streams typed attribute buffers directly to the GPU.

Ableton Live projects reference their samples by absolute path. Across fifteen years, six drives and several machines, most of mine no longer resolved. Scripts catalogue every project, measure which sample references resolve, and export the recoverable sets without modifying an original.

Each drive is walked into a flat index of project files and audio files with their sizes. A master catalogue holds one row per .als copy on every drive, 87,736 rows, deduplicated by name and size into 10,594 unique sets. A schema census classifies each set’s file-reference format: Live 9-era sets store paths as RelativePathElement children beside an alias blob, while modern sets carry Path attributes, and the parsers read both.

For every sample reference in every set, health_scan.py classifies it as resolving project-relative, resolving absolute, recoverable (a file of the same name and size exists on an indexed drive), or gone. On the main archive drive that is 2.03 million references across 7,858 sets. A set whose every reference is resolvable or recoverable is marked fully healable.

heal_export.py plans, fetches, applies and verifies. Every unique sample the healable sets need crosses the network once into a staging pool of 121 GB. Each project folder then receives its samples as APFS copy-on-write clones; references are rewritten project-relative; and the project’s info directory is written alongside. Every placed file is byte-verified against its source, and no original is modified. 3,345 projects were exported from the main archive and 466 from a 2019 backup drive.

For a single set, reconnect.py copies its recoverable samples into the project’s own Samples/Imported/ folder, so Live’s Locate step relinks them itself and no project XML is touched.

One of the six drives failed partway through, its APFS space manager destroyed. It was read without ever being mounted, through libfsapfs; 3,837 files were pulled by path or by inode and size-verified, and the loss among reachable files was measured at 0.05 percent.

A set counts as healed when it opens in Live with no missing-media dialog. The playable share of the main archive went from 8.5 to 51 percent.

A user uploads a photograph of a room, picks a design style and a room type, and gets back a redesigned photograph. The image is encoded client-side and posted to a serverless function that calls a Gemini image model with a long, specific prompt: preserve camera angle, room footprint, and wall, window and door positions, while replacing every finish, fixture and furnishing to match the chosen aesthetic.

Netlify ends synchronous functions at 26 seconds, so generation runs against a fixed internal budget: most of it on gemini-2.5-flash-image, falling back to gemini-3-pro-image-preview with what remains. Credits are deducted with an optimistic-locking update guarded on the user’s current balance, and refunded from a finally block if generation fails.

Anonymous users get three free generations; creating an account adds ten credits. Paid tiers run through Stripe Checkout with a webhook granting credits on completion. The free path has two daily caps: one per identity, and a site-wide circuit breaker on zero-credit generations that bounds total daily model spend.

Supabase provides Postgres, magic-link auth, and row-level security.

Given a seed and a parameter set, the program generates an abstract graphic musical score as vector data, writes it as a clean SVG and a grayscale raster, then runs the raster through a fixed-order degradation pipeline and writes the result as PNG. A command-line tool and a native AppKit GUI are then built on the same core library.

Score generation lays out a series of five-line staff systems, each line shaped by a sine term scaled by a chaos parameter plus per-line random drift, with an independent probability of breaking off early. Each staff carries a note cluster sized by a density parameter — triangular noteheads with stems, occasional articulation dots and connectors — plus annotation marks: rests, arcs, brackets, boxes, arrows, circles, rehearsal marks, numbers, durations, instructions.

The degradation pipeline runs in a fixed order: nearest-neighbor downsample, box blur, per-pixel noise from a deterministic integer hash of coordinates and seed, threshold (flat, or spatially varying if threshold-variation is nonzero), erosion, dilation, then a loop of photocopy generations each re-applying blur, noise, and threshold, and finally a nearest-neighbor upscale back to full size, which is what produces the blocky stair-stepped edges. Because blur precedes threshold, gray edges resolve into hard black or white regions: nearby lines fuse, thin stems vanish, and noteheads merge into staffs.

The program has no third-party dependencies. PNG output comes from a hand-written encoder that wraps stored, uncompressed DEFLATE blocks in a zlib stream with an Adler-32 checksum, inside CRC’d chunks, so files come out at roughly raw bitmap size.

A local evaluation instrument that tests whether an agent can accurately predict how external evaluators will score its work before judgment occurs. For each task, the harness records the agent’s action trace, prompts the agent to forecast its own rubric scores and failure modes, and runs independent LLM judges alongside deterministic reference checks.

An optional intervention pass feeds the agent’s self-audit back into a linked revision run so calibration and self-correction can be measured separately. Every run emits a self-contained directory bundle (manifest.json, trace logs, per-judge verdicts, and disagreement matrices) plus a single-file HTML report.

Eleven self-contained HTML documents, each a personal homepage in the style of the 1998–99 web, generated in December 2025 with Claude Opus 4.5 and framed by a short written statement. The personas include a goth teenage poet, a self-styled hacker, a Y2K survivalist, an anime fan, an amateur photographer, a professor, a New Age contactee, a nu-metal band’s singer, a fictional federal bureau, a new religious movement, and a net.art practitioner.

The construction is period-accurate in markup and static in behavior. Pages use nested table layouts, font tags, marquees, inline hex colors, and a CSS keyframe animation standing in for the blink tag, which no longer renders. Several pages render their animated-GIF slots as bracketed text descriptions. The only script is a snippet on the photographer’s page that disables right-click with a copyright alert, a period-authentic gesture.

The era’s social machinery is present as surface: guestbook forms and visitor counters whose numbers are written into the markup. The net.art page goes furthest into fiction, describing works the persona made — a domain serving procedurally generated 404 poems, a piece rendering the cursor traces of previous visitors — that exist only as that description.

The statement argues that a personal homepage of that period was an act of self-construction addressed to an unknown reader, and that guestbooks, counters, and webrings all functioned as proof that someone had arrived. It holds that these particular personas are fabrications rather than reconstructions of real people, and asks what separates a fabricated persona from a real person performing a version of themselves for an audience — a question it declines to resolve.

A browser tool that loads an image and re-encodes it as a lossy JPEG repeatedly, so that generational compression loss accumulates in the pixels. Each pass calls canvas.toDataURL('image/jpeg', quality), loads the result back into the canvas, and encodes again, so every generation passes through the browser’s own JPEG codec. Defaults are 50 iterations at quality 0.30, adjustable to 200 iterations and any quality from 0 to 1.

Three optional passes can run between re-encodes, in fixed order: per-pixel noise built from a shared luminance term plus a smaller per-channel term, a color shift pushing red up and blue down with a slight contrast adjustment, and a five-point sharpening convolution over each pixel and its four neighbors. All three are hand-written loops over the pixel array, and all default to off. The interface yields to the event loop each iteration so progress can render. Saving exports at quality 1.0, keeping the final generation intact.

The whole application is one HTML file with inline CSS and JavaScript, and images stay in the browser.

An inspection workbench for indexing and comparing algorithmic vision specimens across structural, operational, and epistemic axes. Each specimen is analyzed through a server-side vision pass that extracts formal traits, failure signatures, and curatorial metadata into a typed schema, paired with dense semantic embeddings for spatial projection.

The frontend pairs a faceted taxonomic matrix with a two-dimensional vector canvas. Users can project specimens by cosine similarity, inspect nearest-neighbor clusters, and query the corpus for structural parallels across disparate imaging pipelines.

A command-line archiver that enumerates every workspace and track in a Suno account, downloads the MP3 audio and cover image, and writes per-track metadata (prompt, model version, tags, timestamps, and generation IDs) alongside a master library_index.json.

Authentication reuses an active browser session via Clerk token exchange or an explicit cookie, avoiding manual per-track downloads. Passing --last-run restricts subsequent runs to tracks created after the previous sync timestamp.

A workbench that deconstructs interface screenshots into structured design system specifications. An Express backend submits uploaded captures to a structured multimodal extraction pass that infers color palettes, contrast ratios, modular type scales, spacing grids, and component anatomy.

The interactive workspace renders the extracted system against live preview components and exports W3C Design Tokens Community Group (DTCG) JSON, CSS custom properties, and component scaffolds.

I built a six-stage pipeline for curating and packaging a licensed digicam dataset for a style LoRA, and decided to open source the workflow alongside the dataset. krea-2-trainer. <The pipeline reads and writes through files at each stage, ensuring reproducibility and modularity. Each stage is standalone, so it can be rerun independently or modularly combined with others. The final output is a zip file ready for training, accompanied by caption sidecars for each image and an attribution CSV. >

scrape.py harvests 640px thumbnails from Wikimedia Commons “Taken with…” categories for eleven cameras — Nokia N95, Sony Ericsson Aino, C902, C905, and C702, Canon PowerShot A80, A70, and A95, Sony DSC-P200 and DSC-P8, Kodak EasyShare C743 — capped at 1,000 per category. A file is kept only if its license is CC0, public domain, CC BY, or CC BY-SA. An EXIF date outside 2003–2010 is rejected; a missing date is kept and flagged. Each kept file is a row in manifest.csv. Other fun usecases for this workflow include curating datasets for different camera types, eras, or licensing constraints, and experimenting with alternative ranking and captioning models.

rank.py scores those thumbnails with SigLIP 2: mean cosine similarity to six candid, flash-lit, party-scene prompts minus mean similarity to six encyclopedic landmark prompts. The score is a sort order, written to scores.csv.

gallery.py writes one static HTML file, sorted by that score. A click cycles keep, reject, unmarked. Export downloads keepers.json. The committed cuts are 160 combined, 122 cameraphone, and 38 compact digicam. This involves manually reviewing and curating the images to ensure they meet the desired criteria for the dataset, but the workflow is designed to make this process as structured and reproducible as possible and it is automatable [link].

fetch.py downloads the original for each keeper into data/fullres/, skipping files already on disk.

caption.py asks MiniCPM-V, the model chosen at runtime through Ollama, for the literal content and forbids style, camera, and era. A caption that is too short, or one word repeated past a fixed count and fraction, is rejected. This is done via a loop that continues until a valid caption is produced or the model fails. The loop ensures that only acceptable captions are retained. A rejected or unreachable reply falls back to Qwen3-VL, which sees a 1024×1024 copy, reloads for every image, and starts each reply with one of six openers chosen from the filename. Sidecars are written as content plus a trigger. The default trigger is y2k digicam snapshot style. The reason Qwen3-VL is used as a fallback and not the primary model is to ensure is a matter of speed: MiniCPM-V is faster for the majority of cases, and Qwen3-VL is only invoked when the primary model fails to produce a satisfactory caption. Your own environment can be configured similarly, choosing the primary and fallback models based on performance and reliability considerations.

package.py will not write the zip unless every keeper has an image and a caption sidecar. Beside the zip it writes an attribution CSV. Passing a keepers file and an output name packages one cut from the same full-resolution pool.

Depending on where you end up training your model, you may need to create a trigger that matches the expected input format for your training pipeline. This could involve renaming the default trigger, adjusting the sidecar structure, or modifying the caption content to align with your model’s requirements. The current default trigger is y2k digicam snapshot style. There is a chance this actually leads the model to focus too narrowly on the style descriptor as a content vector rather than style itself, so careful evaluation and potential adjustment of the trigger may be necessary.