Two ventures of my own with two different roles: one platform live and in use, one institutional platform in development. What connects them is how they are built.
Language & learning platform for students
A modern language and learning platform for students, built by me and live on the web. Made with the same agentic method as my regulated AI work: a multi-agent workflow of parallel streams, loops and skills.
Built with
The NLP and audio stack behind the live platform, public on GitHub: razbiram-nlp · razbiram-listen
View liveA syllabus-guided AI study coach for universities
Sovereignty, isolation, evidence and an audit trail, built on my own production system first.
Problem
Students — especially in medicine — drown in fragmented curricula; the cross-links between subjects stay invisible, and current research is decoupled from the actual learning path.
Built
A swarm of cooperating agents manages the knowledge around a course of study across many nodes: it couples a university's syllabus dynamically to each student's guided learning path, weaves in current research, and generates several didactical reasoning pathways — medicine as the pilot. Underneath, a sovereign, auditable knowledge pipeline already runs (below) — moving step by step toward a medical reasoning system.
Value for a client
A university couples its syllabus, alive, to each student's individual learning path and connects study with research — explainable, evidence-based, with traceable reasoning pathways rather than a mere knowledge store — built on the same sovereignty and auditability principles as regulated AI.
The agentic knowledge pipeline (already running)
Deterministic scan: what’s missing is derived from the files, never from a stored snapshot.
aya-expanse:8b via Ollama, temperature 0, SQLite checkpoint. No API or token costs, nothing leaves the device.
Every Cyrillic token must survive verbatim, otherwise the translation is rejected.
Surgical JSON patch with a JSON.parse guard: the file is never written invalid, preventing data corruption on system crashes.
Only hard cases go to a hosted model (Haiku), and through the same gate.
Machine-checked gates and an append-only ledger per wave; commit only when all is green.
Local & deterministic first · a hosted model only for hard cases · commit only when every gate is green.
Principle → implementation
Sovereignty
Translation runs locally (aya-expanse:8b via Ollama, on-device). Bulgarian content never leaves the machine, and there are no API or token costs.
Isolation
API, app and data are separated (Docker); the API container deliberately cannot read the content directory. Secrets are checked via gitleaks in CI.
Evidence
Every AI call is logged (model, tokens, latency). Knowledge-graph edges carry source, confidence and soft-delete. AI edges are treated as low-trust.
Audit trail
AI features are individually switchable, off by default; every admin action and tier change is appended to the audit log, append-only, enforced by a DB trigger.
The same architecture — locally deterministic, audit-trail before anything else — is documented publicly in local-agent-pipeline.
Own project in development, not a finished client reference. Architecture and figures come from my own production system.
Architecture note (PDF)Two services, one way in: a 20-minute introductory call. What fits your venture is set out in the offering.
View the offeringThree workpieces. Three clocks of their own.
Only when they come apart can you see what carries them.
Put back together, the decisions disappear into what works.
