EntroπaLabs
Computer vision · Language models · Robotics

The world tends toward disorder. We build systems that hold.

Entroπa Labs is an independent AI practice. We take perception, language and robotics from a result that holds on a benchmark to one that holds in the field — and we publish what we learn along the way.

Enter the playground Read the research

Practice

Three fields, one discipline: make it hold outside the lab.

Benchmarks reward the average case. The world serves the tail — drift, noise, edge conditions. We work where models meet reality, and measure what survives contact.

Perception

Computer vision

Perception that degrades gracefully — 3D reconstruction, detection and tracking tuned for the conditions a demo never shows you.

Language

Language models

LLM systems you can inspect and trust — routing, evaluation, and the plumbing that turns a capable model into a dependable one.

Embodiment

Robotics

Control and planning that close the loop under uncertainty — policies that recover when the world refuses to cooperate.

Writing

One post per system, from the source.

Not announcements — explanations. Each one takes a single application apart: the mechanism underneath, the part that turned out to be harder than it looked, and the limitations left in rather than tidied away.

7 postsNewest first
every one open to inspection

26 Jul 2026The agent loop, and what it can't do in a browserA tool-use agent is about forty lines of control flow. Everything difficult about it lives in the tools you are willing to hand it.Agentic AI12 min25 Jul 2026Reading a checkpoint in the browsersafetensors is a length, some JSON and a wall of bytes. A torch .pt is a ZIP with a pickle in it — which means writing a small pickle VM.Generative AI13 min25 Jul 2026Deciding which experts to dropMost of an MoE's experts are underused and several are near-duplicates. Turning that into a defensible pruning plan needs two numbers, not one.Generative AI11 min25 Jul 2026A 3D renderer in 120 lines of canvasNo WebGL, no matrices library, no dependencies: project, cull, shade, sort. The robot playground draws every frame the way graphics worked before GPUs.Physical AI11 min25 Jul 2026A forecast that grades itselfAnyone can publish probabilities. The interesting question is whether they mean anything — so this one scores itself every night and refits.Sports14 min25 Jul 2026Valuing a hockey player from a box scorePoints tell you who is good. Surplus value tells you who is *tradeable* — and the gap between those two is where a GM actually works.Sports13 min24 Jul 2026How the Gaussian splatting viewer worksA radiance field is millions of fuzzy ellipsoids. Getting them on screen at 60fps is mostly a sorting problem, and the sort is the interesting part.Physical AI12 min
All writing

Our research

Seven working systems, each one open to inspection.

Our research ships as software you can run, not screenshots. Each entry below is a live application — the purpose it was built for, the mechanism underneath, and the door into it. Everything executes locally in your browser: no file you open ever leaves your machine.

7 / 7 liveAcross four practices
generative · agentic
physical · sports

Generative AI

The models themselves — opening a checkpoint, reading the weights, deciding which experts earn their parameters.

2 systems
Agentic AI

Systems that decide and act unattended — a tool-use loop that picks its own tools and runs them against live data.

1 system
Physical AI

Systems tied to the physical world — reconstructing a captured space, and driving a body through a simulated one.

2 systems
Sports

The sports models — a hit forecast that grades itself against last night's box scores, and a salary-cap engine that values every skater and trade.

2 systems
Project Orion Physical AIPerception · 3D reconstruction Live

Gaussian Splatting Viewer

PurposeJudge a radiance-field reconstruction by moving through it. A 3DGS scene looks flawless from the training views and falls apart two metres to the left; this puts free viewpoint control in front of the reconstruction so the failure shows itself.

Function
  • Loads a .splat or an Inria .ply straight off disk
  • Packs every gaussian's covariance into an integer data texture, renders with instanced billboards
  • Depth-sorts every frame in a Web Worker so the main thread stays free
  • Orbit, zoom, shift-pan, and a splat-size slider; a demo scene is embedded so it is never empty
WebGL2Web Worker sort.splat / .plyno upload How it works →Open the viewer
Project Cygnus Generative AILanguage models · Interpretability Live

Weights Visualiser

PurposeMake a checkpoint legible. A model file is usually a black box you load and hope about. This one opens it in the browser and shows the architecture, the actual weight values, and where the tokens go.

Function
  • Parses .safetensors in full; torch .pt/.pth best-effort via a ZIP reader and a minimal pickle VM
  • Architecture chart of parameters per layer, log-scaled and typed by colour
  • Weight matrix as a diverging heatmap, plus a value histogram with min/max/mean/std/%≈0
  • Token routing — a simulated forward pass animating each token to its top-k experts with live per-expert load
safetensorstorch pickle VMMoE routing simno upload How it works →Open the visualiser
Project Lyra Generative AILanguage models · Efficiency Live

Mixture-of-Experts Pruner

PurposeDecide which experts you can drop, and say what it costs. Most of an MoE's experts are not pulling their weight and several are near-duplicates of each other. This makes that argument with numbers instead of intuition.

Function
  • Utilisation — share of routed tokens per expert, from a seeded router simulation
  • Redundancy — cosine similarity between down-sampled per-expert weight signatures
  • Keep-score ranking: normalised utilisation minus a redundancy penalty
  • A live pruning plan — experts cut, parameters saved, and a clearly labelled heuristic retention estimate; downloadable as a standalone report
router simulationcosine redundancyclient-side reportno upload How it works →Open the pruner
Project Pegasus Physical AIRobotics · Embodiment Live

Robot Playground

PurposeBuild intuition for embodied control with nothing installed. Morphology and terrain decide as much as the controller does; here you can change both in a second and feel the difference on the keyboard.

Function
  • Five bodies — humanoid, quadruped, rover, robot arm, drone — built from nested articulated frames
  • Five worlds — studio, grid arena, obstacles, ramps, low-gravity moon
  • Procedural gaits driven live: WASD to move, space to jump, R/F for altitude, per-joint keys for the arm
  • A dependency-free software 3D renderer on a 2D canvas: project, backface-cull, flat-shade, painter's sort
software rasteriser5 bodies × 5 worldszero dependencies How it works →Open the sandbox
Project Aquila SportsApplied forecasting · Calibration Live · daily

Beat the Streak

PurposeRun a probability model in public and grade it every day. Calibration is the only claim that matters and the only one most models never publish. Baseball supplies a fresh, honest test slate every morning.

Function
  • Ranks every likely starter by modelled P(at least one hit) for today's slate
  • Features: matchup, contact and plate skills, recent form, platoon split, ballpark, batter-vs-pitcher history
  • Rebuilt each morning by a stdlib-only Python pipeline on a scheduled workflow
  • Yesterday's board is graded against results and kept as history, so the weights answer for themselves
daily refreshself-gradingstdlib-only pipelineMLB API How it works →Open the board
Project Draco SportsDecision support · Valuation Live · daily

NHL GM Playbook

PurposeTurn public data into the decision a general manager actually faces. Not "who is good" but who is underpriced, who is movable, what a team is missing, and which side wins the trade.

Function
  • Scout — breakout score, box-score WAR and $/WAR, and surplus trade value for every qualified skater
  • Contract projections — next-deal AAV and term, live from CapWages with runtime anchor validation and a CSV fallback
  • Team Lab, all 32 clubs — optimal lines and handedness-balanced pairs, starter, roster needs, movable players, farm pipeline, fair trade ideas
  • Trade analyzer — build both sides and get a value verdict on who wins
NHL APICapWages contracts32-team labtrade analyzer How it works →Open the playbook
Project Vega Agentic AILanguage models · Agents Live

Voice Agent

PurposePut the agent loop somewhere you can watch it work. Tool use is usually described in a diagram; here you ask a question out loud and see the model choose a tool, the tool run against real data, and the answer come back — the whole cycle, in the open.

Function
  • Speech in, speech out — Web Speech recognition with a live level meter, answers spoken back; typing works everywhere
  • Two provider adapters, the only two wire formats there are: Anthropic content blocks and the OpenAI-compatible tool-calls array
  • Tools query this site's own live data — today's hit probabilities, skater valuations, a club's roster needs
  • Your key, your browser: kept in localStorage and sent only to the provider — there is no server here to send it to
Web Speech APItool-use loopBYO keyno backend How it works →Open the agent

Each system has a written companion on the blog — what it is for, the mechanism underneath, and the part that was harder than it looked. Papers land here as the work produces them; the shelf is deliberately empty rather than padded. The systems above are the current record: reproducible by construction, offline by default, and honest about which figures are measured and which are estimates.

Contact

Have a system that has to hold in the field?

[email protected]

Independent AI practice · computer vision, language models and robotics · assessments and hands-on engagements.