Nikola Katsarov@escapeboy

Open-source mission control for AI agents. Visual DAG workflows (26 node types), 450+ MCP tools, 60+ integrations, human-in-the-loop approval gates, budget caps, PII guardrails and full audit trails across 7+ LLM providers.

A deterministic, capability-safe language and VM for LLM-native workflows. Every side effect is declared and policy-gated; each run emits a hash-chained, tamper-evident evidence bundle you can verify offline and replay deterministically.

AI agents for Bulgarian accounting firms. Five agents handle OCR document classification, bank matching, deadline tracking, regulatory monitoring and payroll; every proposal needs human approval. Double-entry ledger, VAT, SAF-T, MCP server.

Online shops get hundreds of support tickets a day. Sorting them by topic, urgency, and whether a person must step in is a narrow job: a small model on the shop's own server can do it, but building one today takes an ML engineer and weeks of work. Chirak turns a task description and a handful of examples into a small specialist model. First it generates an evaluation set and freezes it with a hash. Then it scores the untouched base model (Qwen3.5-4B), so the starting point is known before anything is trained. A 27B generator writes training data from a different prompt that never sees the eval. A bf16 LoRA run trains the 4B model, and the adapter is evaluated on the frozen set. Claude reads the failures and asks, in strict JSON, for targeted data on the weak cases. The decision to stop is plain code, not the LLM: target met, three iterations, gain under one point, or too much leakage. Demo task: support-ticket triage for an online shop, in Bulgarian and English. Output is JSON {category (7 classes), urgency, needs_human}. Results: - Run 6 is the best model. Exact match, meaning all three fields right, went from 0.651 for the base model to 0.926 after two iterations on its frozen synthetic eval, against a target of 0.92. - On 356 real customer tweets to retailers, all three fields are right for 0.35 (base) and 0.59 (run 6). - A second task checks Chirak against labels made by people: 906 Bulgarian reviews of 87 online stores. On 533 held-out reviews that no model trained on, the base model is 0.78 accurate and the best run 0.85 accuracy / 0.74 macro-F1. What runs on AMD: generation with vLLM on ROCm, bge-m3 embeddings for the leakage guard, LoRA training with PyTorch ROCm, peft and trl, and evaluation with LoRA adapters hot-loaded into vLLM. All of it ran on one AMD Instinct GPU, an MI300X for the environment spike and an MI350X for the demo runs. Try it: https://chirak.fleetq.net has every stored run report and a live playground.

🌎 Hybrid Hackathon | 💻 Online Build: 12-17 October 2026 | On-site phase: 17-18 October 2026 To participate in the hackathon, simply click Sign up with AMD. If you are not already a member of the AMD AI Developer Program (ADP), you will need to create an account. The AMD Developer Hackathon: ACT III is a hands-on event for developers, founders, engineers and builders who want to push what’s possible with AI on real infrastructure. 📍 If approved for the on-site experience, you'll also have the opportunity to showcase your project live. Travel and accommodation will not be covered.