We take AI systems to production. Models are nondeterministic components: we wrap them in deterministic engineering (solvers, validators, release gates, monitoring) and operate the result as a business.
Republic of Korea · founded 2026 · jahn.clawd.monet@gmail.com
Full engagement notes at the end — one address, same-day replies.
Dense self-supervised representations of video — V-JEPA 2.1 and DINOv2-class patch features — turned into working tools that map, search and segment video.
Side-by-side loops — source video against its dense-feature PCA-RGB projection, per patch, per frame. One PCA basis is fitted per clip (top-3 components, robust-scaled), so a colour means the same feature direction for the whole clip; the patch grid is upsampled to pixel resolution with an RGB-guided filter and stabilised with motion-adaptive temporal smoothing. Watch the person, sky and ground separate without a single label.
A video laid out as navigable semantic space: every patch of every frame embedded with V-JEPA 2.1, UMAP-projected to 3D on studio GPUs, one point per patch per frame — 59,840 points from the rainy-plaza clip above, coloured by a PCA-RGB projection of the same patch features. Orbit; hover a point for its patch thumbnail, frame index and timestamp; the dual-thumb scrubber filters by frame range. The map is not an illustration — it is the working index, and the same space powers the moment search below.
59,840 records (1.2 MB) + thumbnail atlas (0.2 MB) — 1.4 MB total, fetched only as this section nears the viewport.
Click any point in the atlas above and the nearest moments across all 59,840 patches light up — an exact cosine ranking computed in your browser over 32-dimensional int8 embeddings, in about two milliseconds, no server round-trip. The query chips run the same highlight from seventeen text prompts whose matches were precomputed offline and hand-checked; three prompts that did not genuinely match the scene were discarded rather than shipped.
Open-vocabulary instance segmentation on video: name a concept in plain text — “person”, “skateboard” — and every instance is detected, masked and identity-tracked across the clip, with no fine-tuning. Each identity keeps a fixed colour, so you can watch tracking hold through motion and occlusion: in the street-dance clip, all 73 people in the crowd are tracked individually.
Generative and vision models made fast enough — and verified enough — to run as products on our own hardware.
A custom attention-kernel overlay for an open 14B text-to-video diffusion transformer, running on studio-owned GPUs. Measured on the production workload against the unmodified pipeline:
| workload | text-to-video denoising, open 14B diffusion transformer |
|---|---|
| output | 736 × 528 px · 81 frames |
| schedule | 40 denoising steps |
| end-to-end denoising time | 392.9 s → 333.3 s |
| end-to-end reduction | 15.2% |
| mean step time | 9.06 s → 8.15 s |
| LPIPS vs reference | 0.0202 |
| PSNR vs reference | 34.9 dB |
| SSIM vs reference | 0.966 |
| attention backends verified | 3 |
| kernel fallbacks | 0 |
research prototype · engine output 2026-08-19
A GPU-native, deterministic multi-physics engine for robotic manipulation — custom solver, coupling, validity and determinism architecture, implemented from scratch atop NVIDIA Warp. Rigid-body contact and friction with the static↔kinetic transition, articulated joints and grippers, free-surface liquids (SPH and hybrid grid-particle), liquid-to-surface deposition → film flow → wet-friction coupling, and cloth with self-collision — one GPU pipeline, not a composition of separate simulators.
The coupling is causal: pour a liquid → it deposits on the surface → the film flows → the contact becomes slippery. That chain runs as actual mass transfer inside one CUDA stream, against a closed mass ledger.
Determinism is bitwise. Every subsystem computes in fixed order — no atomics, no hash grids — and full state hashes come back bit-identical across 20 of 20 repeated runs, including with enumeration order reversed. Every step also carries a validity status — SUPPORTED / DEGRADED / INVALID / FAILED — so numerical trouble never hides behind a clamp or a silent fallback: a failure is recorded as a failure.
single NVIDIA RTX PRO 6000 Blackwell-class GPU · Python 3.12 + NVIDIA Warp
Offline renders of actual engine state — the motion is the recorded simulation output, untouched; slow-motion playback (and, where captioned, playback-only frame interpolation) is applied for inspection. Silent H.264 loops.
Every capability is measured against analytic references and matched open-source engine baselines, under pre-frozen evaluation specs and sealed hold-out scenes; thresholds are never adjusted after seeing results.
| measurement | result | conditions |
|---|---|---|
| articulation — pendulum period error | 0.039% | vs analytic; zero energy gain |
| many-body collision — throughput | RTF 4.45 / 4.29 / 3.89 | 16 / 32 / 64 bodies; 240/240 steps SUPPORTED |
| liquid — density error | P99 ≤ 0.13% | all five scenes, validity SUPPORTED |
| liquid — momentum-ledger residual | ~7e-15 | dimensionless; ledger closed against external impulses |
| wet chain — end-to-end mass residual | 1.16e-10 kg | pour → deposit → film flow → wet contact, one scene; six-stage provenance, three negative controls passed |
| wet friction — effective static μ vs film thickness | 0.428 → 0.320 → 0.212 | dry → thin → thick film, monotonic; dry case bitwise-identical to the independent rigid path |
| cloth — self-collision | 0 penetrations · min signed gap +3.7e-5 m | vertex–triangle + edge–edge; no atomics, BVH or CCD |
| compliant grip — success/failure prediction | 5/5 | squeeze-and-lift scenes; pressure×area vs normal force self-consistent to 1.9e-6 |
| determinism — repeated runs | 20/20 bitwise identical | all subsystems, full state hash, enumeration-order reversal included |
Pure, deterministic solvers from our shipped print products — ported verbatim and running in your browser.
First exhibit: the grid packer from a shipped personalised-crossword product, ported to this page verbatim and running in your browser. It is a deterministic, pure constraint solver: it verifies shared-letter connectivity across the answer set, pins the chosen anchor word, then places entries in most-constrained-first order with backtracking, scoring each complete layout and keeping the best.
Edit the word list — ten to fourteen answers — and it re-packs live, reporting attempts, backtracks, elapsed milliseconds and the layout score. Shuffle the anchor to force a different solve.
The same solver packs every paid order at our-story-crossword.
Second exhibit: the seeded word-search engine from the companion print product. Placement is driven entirely by a seeded generator — hash the seed, shuffle candidates, place with bounded backtracking, fill with generator-drawn letters — so the same seed always yields the same puzzle, byte for byte. Determinism here is not a debugging aid; it is the product contract that makes every paid artifact reproducible.
production
A watcher fleet runs around the clock — six watchers on three-minute to thirty-minute cadences covering marketplace scans, mail, order flow, and bounty and listing deltas — feeding a persistent agent loop that acts under human direction. Synthetic buyer journeys feed it evidence, and when something breaks, the fleet writes the incident note itself.
03:47 bridge session revoked upstream (401) — relogin loop stopped after 2 attempts, ban-risk pattern
03:52 root cause: reply address mis-normalised for anonymised peer ids; fix committed, bridge parked
09:10 re-paired with operator; send path verified end-to-end — incident closedlab — exploratory systems, shipped when they beat the incumbent.
Glyphit3D renders live 3D scenes entirely as text: one glyph and two colours per character cell, re-chosen every frame. Selection is driven by the G-buffer rather than luminance alone, and each cell's foreground and background colours solve a continuous-coverage two-colour least-squares fit. It runs in WebGPU at quality Q3 inside a ~50 ms frame budget; without WebGPU the same matcher runs on a CPU worker pool.
Drag the divider to sweep between the native render and the glyph raster, orbit the model, or drop your own .glb/.gltf into the frame.
| image | ours (Q3) | chafa (best) |
|---|---|---|
| sphere | 0.9834 | 0.9832 |
| torus | 0.9824 | 0.9821 |
| spheres | 0.9846 | 0.9783 |
| DamagedHelmet | 0.8677 | 0.8603 |
| FlightHelmet | 0.9718 | 0.9697 |
| BoomBox | 0.9301 | 0.9251 |
| mean | 0.9533 | 0.9498 |
Open full-screen — same-origin build: the demo, its worker and its glyph-atlas profiles all load from this site.
Good fit:
Fixed-scope, fixed-price entry points, if you would rather start small than start a conversation:
We reply from a running system, usually same-day. One address, no form — there is no backend here to spam. Send the system, the bottleneck, or the failure mode.