The Lab
Most consultancies show you screenshots. These are running right now.
Five working pieces of the systems we build. Two run in your browser as you scroll, two replay real recorded runs, and one generates a firewall policy from a behavioural baseline. No video, no mockups.
A matching engine that shows its work
The production version scores résumés against 400,000+ job postings. Every point it awards carries a reason code, and it refuses to recommend anything the evidence doesn't support — it cannot claim a skill you can't back. This is that engine, scaled down to our own service lines.
Deterministic scoring · no model · fully auditable
Step 1 — describe the situation
Select what's true for you, then run the engine. It will show every point it awards.
A swarm claiming work off a shared board
Agents claim subtasks, open channels to peers, and converge until the board is empty. This is the shape of the orchestration layer behind a multi-agent engine refined across roughly 180 build sessions — the unglamorous part that separates a working system from a demo.
Seeded PRNG · reproducible runs · canvas
message bus
press run to start the swarm
What LoRA actually costs you
Before anyone fine-tunes anything, the question is what it costs. Move the rank and watch trainable parameters and training memory move with it. The formula is printed next to the answer, because you should be able to check it.
Live arithmetic · real model geometries
base model
adapted matrices
trainable = r · (d_in + d_out) · matrices · layers
square attention projections ⇒ 2·r·d per matrix
Ten agents arguing until they agree
A real job, replayed from its own event log: Claude, Codex, Gemini and Kimi working the same specification through eight phases. Adversarial critics fail the work and hand it back, three auditors review in parallel, and a coherence pass reconciles them round after round. Watch the gate ledger rather than the floor — this run ends with the technical critic still refusing to sign off after three attempts, which is what a review system looks like when it is actually reviewing.
Recorded run · 302 minutes · auditors converged on round 3 of 4 · one critic never passed
phases
convergence
not yet evaluated
gate verdicts
- none yet
One completed job, replayed from its own event log — 179 state changes over 302 minutes, compressed. Ten agents: adversarial critics paired with revisers, three auditors running in parallel, and a coherence pass that reconciles them. The auditors converged on round 3 of 4, and the ledger ends on three failures, because the technical critic never accepted the work and the round budget ran out. That is the real result, not a tidied one — an orchestrator that always agrees with itself is not reviewing anything. The engine never leaves the machine that ran it; this page receives state changes only.
A firewall policy written by watching
Servers get outbound rules by hand, or not at all, so a compromised process can talk to anywhere it likes. This agent scores every destination a workload reaches by how much evidence supports it, then renders an nftables policy where default is deny and only well-evidenced destinations become live rules. It loads with a real capture from an actual machine — where it enforces nothing, because a process name is not a stable identity and two minutes is not a baseline. Switch to the 14-day journal to see a mature policy, and drag the observation clock to watch rules fall out of enforcement as the evidence goes stale.
Engine runs in your browser · real capture + synthetic baseline · eBPF collection specified, not built
stage 3 — evidence and confidence
Stages 2–4 of the agent, executing in this tab as you move the controls — the journal is folded into per-workload profiles, every destination is scored against the specification's confidence rules, and an nftables policy is rendered where HIGH becomes a live rule, MEDIUM ships commented out pending review, and LOW is never emitted. Stage 1 collects events with eBPF hooks in the kernel; that part is specified, not built, so the real capture here was taken by polling the OS connection table instead.
One more, on the homepage
And a search engine that runs on your machine.
An embedding model loads into your tab and answers questions about this studio by meaning rather than keywords — with a confidence floor, so it tells you when it doesn't know.
Try it on the homepageRelated
semantic rankingContinue with another relevant resource.
- Related resourcecosine match 0.70Selected work: built, tested, measuredClosest content match: Selected work: built, tested, measured.Open resource
- Related resourcecosine match 0.64About the studioClosest content match: About the studio.Open resource
- Related resourcecosine match 0.57Zero-cost AI playgroundClosest content match: Zero-cost AI playground.Open resource
Want one of these pointed at your problem?
These are demos. The versions we build for clients run over real data, at real scale, with the evaluation and guardrails that make them trustworthy.