Local AI laboratory · Apple Silicon
Karaibart builds AI systems that run entirely on your own hardware — and proves it, numbers in hand. Multi-model inference infrastructure, sourced RAG, tool-using agents hardened against the OWASP ASI Top 10.
Simulation: request routing, 80 GB RAM budget, LRU eviction. Click an organ.
The founding insight
I designed and shipped a fully local AI ecosystem: eighteen specialized applications — document RAG, coding agent, speech synthesis, music generation, symbolic score analysis, lore engine, sample indexer, memory guardrail.
Each loaded its own models and ran its own memory — roughly 208 GB of redundant weights on a 128 GB machine. The architect's call was not to write yet another application, but to build the infrastructure that multiplies the value of everything already there: a RAM-aware gateway, a single budget, an eviction policy that decides.
Since then the ecosystem has cleared three milestones: full independence (the last third-party runtime is gone), agentic hardening (OWASP ASI Top 10 translated into real controls, supply chain taken from 122 CVEs to 0) and proactivity (the assistant observes its own usage and proposes, under a hard anti-spam cap).
Zero cloud, zero telemetry, enforced loopback bind. Your data never leaves the machine.
A RAM-aware gateway shares tens-of-GB models across every application under a single budget.
Supervised services, nightly quality eval, reusable abstractions. Engineering, not prototype.
Every claim is measured before it is written: phys_footprint rather than ps rss, cosine verified before migration, latencies quoted cold as well as warm.
The laboratory
Every project has its page: what it is, the architect's call, and what went wrong. Post-mortems are worth more than screenshots.
Control plane
RAM-aware MLX gateway (80 GB budget, LRU eviction, pinned model), shared memory bus, ReAct intent router, usage journal and a 10-section nightly eval. The "nervous system" that unifies the whole ecosystem.
Tabular foundation model
SOTA tabular ML, 100% local: calibrated classification + SHAP explainability. 694 books recategorized in the live database, offline medical demo.
100% local RAG
Answers sourced only from the books (zero hallucination). 24,799 books, 1.68 million chunks, a 21 GB database, cross-encoder reranking, cross-lingual.
Security & compliance
An agent that executes code, reads files and remembers is an attack surface, not a demo. These are the controls actually running on my own ecosystem — the same ones I install for a client.
First audit pass across four production environments: 122 known vulnerabilities. Cleared entirely — including the real root cause, a command-line tool installed inside a service's virtual environment, pinning stale versions there. Three CVEs with no upstream fix are tracked by name and reviewed monthly. A nightly audit section now goes red the moment one reappears.
Any model-produced code runs in a sandbox: network cut, writes confined to a throwaway directory, reads of SSH keys, cloud credentials and keychains denied. A probe tests the jail before each run — leaky jail, red build.
ASI05The memory → prompt path is sanitized end to end, agent and speech synthesis included. A note stored yesterday cannot become an instruction today.
ASI06Nine write sinks go through symlink-resistant path resolution bounded to declared roots; model names are filtered by allowlist.
ASI0217 services catalogued, all loopback-bound, a single external credential across the whole ecosystem — and a flag reduces it to read-only.
ASI03Answer quality, code executed under pytest, supply chain, system posture, journal freshness. A red verdict is a signal, not decoration: it has already triggered real fixes.
GatesFirewall on in stealth mode, disk encryption, every service bound to 127.0.0.1, telemetry off by default, secrets never inline in service files.
HostServices
For organizations whose data cannot leave the building: healthcare, legal, defence, finance, public sector. An engagement ladder — start small, continue if it earns its keep.
Security & reliability of an LLM or agentic system
I take your system as it actually runs and look for where it breaks — not in theory, in your code.
Read →On-premise LLM stack, 100% local
The reference architecture behind this portfolio, fitted to your hardware and your regulatory constraints.
Read →Continuous eval, CI gates, observability
An AI system degrades silently. Mine is watched every night; yours can be too.
Read →That is precisely the problem I solve. A 30-minute call is enough to scope an audit.