AI SYSTEMS
THAT HAVE TO
WORK OUTSIDE
THE LAB.

TANS

ProblemA model can benchmark well and still be unusable on the target machine because latency tails, peak RAM, sustained thermals, energy draw or accelerator constraints were treated as an afterthought.

SolutionTANS adapts the model and evaluates the actual target together. It can fine-tune foundation models or detectors for the domain, distill from a larger teacher, prune unnecessary structure, quantize precision, export to the required runtime and measure every change on the device.

OutcomeNot simply “a smaller model”, but a reproducible deployment candidate selected against quality, p95/p99 latency, memory, thermal throttling and power/energy budgets — with changes accepted, rejected or rolled back from measured evidence.

Vigilant Inference Governor

ProblemA robot or vehicle can execute every requested model efficiently and still react late because an old camera frame has no value to the current control loop.

SolutionVigilant Inference QoS sits in front of NVIDIA Triton or TensorFlow Lite and understands age-of-information. It drops superseded work before dispatch, refuses work that would finish too late, protects periodic perception and can select a smaller model variant when only that variant still fits the time budget. vig autotune measures the actual models on the target hardware and can also conclude that no governor is needed.

OutcomeKeep the existing models and inference backend; change which work deserves scarce accelerator time. In the public public Inference Governor demos, including the public Inference Governor workload, the protected camera remains fresh in 100% of cycles instead of 0.1–0.2% under the demonstrated shared-GPU load.

VIDEO / INFERENCE GOVERNORYOUTUBE ↗

External video: this preview is stored locally. No YouTube player, thumbnail or tracking resource is loaded automatically. A connection to Google/YouTube begins only after you click the public demo video.

MEASURED → tuned Triton detector coverage 85% → Vigilant 99% on the measured GPU benchmark · ~20× fewer missed cycles there · four-camera demo: protected stream 0.1–0.2% → 100% fresh cycles under the demonstrated load · tested on NVIDIA GPU and Android/Adreno paths · 900 passing tests in the public repository

Project page ↗ GitHub / evidence ↗ Pilot →

External links: GitHub and GitHub Pages are operated by GitHub, Inc. A connection is established only after you follow the respective link.

Autonomes Reaktives Entscheidungs-System

ProblemConventional security stacks generate detections and alarms, but a control room still has to correlate cameras, thermal feeds, radar and earlier events before deciding whether something is relevant and what should happen next.

SolutionThe system builds a continuously updated situation picture directly at the edge. Domain-specific detectors feed tracked entities and observations into a temporal knowledge graph; source, time, confidence and provenance remain attached. Graph retrieval, rules and reactive decision logic evaluate the evolving situation instead of treating each frame as an isolated event.

OutcomeOn-edge situational awareness can classify and prioritize security events, create a structured local situation report and trigger predefined workflows — for example cue another sensor, notify a control room, or dispatch the nearest available response resource — without requiring cloud inference.

UNSCRIPTED / Social Dynamics Runtime

ProblemMost open worlds have one invisible shared brain: a flag changes and every NPC effectively knows the same state. An LLM can make dialogue richer without fixing that underlying world model.

SolutionUNSCRIPTED is the runtime behind this approach. Each NPC holds local knowledge with source, confidence and history. Information can spread, mutate, be rejected or forgotten; trust and group membership affect belief; locations carry social state. The runtime decides what happened and what a character may know. A language model is optional and can be limited to wording.

OutcomeThe player changes the world dynamically. Witnesses react first, information travels through the social graph, a street or district can become more suspicious or closed, and characters encountered later can behave differently because of what actually reached them — not because a designer scripted that exact dialogue branch.

VIDEO / RUMOUR · LIES · MEMORY · CONSEQUENCEYOUTUBE ↗

External video: this preview is hosted locally. No YouTube player, thumbnail or tracking resource is loaded automatically. Clicking opens YouTube and establishes a connection to Google/YouTube.

External links: the showcase is hosted on GitHub Pages and the source repository on GitHub. They are not embedded here; the connection begins only when you click.

Fused Geolocation

ProblemIn tunnels, urban canyons or under jamming/spoofing, a normal location API may keep returning a coordinate without telling the application whether it should be trusted.

SolutionThe implemented stack fuses local sensor and satellite evidence with a receiver-clock-aware Error-State Kalman Filter, bounded RAIM/FDE-style exclusion and explicit integrity/threat diagnostics.

OutcomeThe consumer receives both the fused position and an explanation of degraded confidence. The architecture is built to continue conservatively when clean GNSS disappears rather than silently treating all fixes as equal.

Vigilant ESP / Experience Semantic Protocol

TheoryHuman language compresses many different dimensions — factual knowledge, intention, emotion, sensory qualities, context and temporal continuity — into a linear sentence. ESP asks whether machines need that linguistic bottleneck at all. It represents those dimensions in typed semantic spaces so that an experience or knowledge state can be encoded as a structured vector state with metadata rather than first being rewritten as prose.

Protocol ideaA sender can transmit a machine-readable semantic state — for example what is known, how something felt, what outcome is desired and with what confidence or provenance. A receiving system can compare, transform or combine that state directly. Natural language, an image, sound or another interface becomes a rendering layer at the edge, not necessarily the transport format itself.

First implementationThe current media engine is the first practical implementation of this post-linguistic idea. A film is mapped into an experience genome; a user's desired experience is mapped into the same semantic space. “Like this film, but darker, faster and less violent” becomes a vectorized state transformation rather than only a keyword query.

DirectionThe longer-term goal is broader than media recommendation: agent-to-agent exchange of experience, knowledge and intention in a structured semantic representation that preserves dimensions which ordinary text can flatten or omit.