# borkiss — building the layer between frontier models & production > Machine-readable mirror of https://borkiss.net (human version is a WebGL/three.js experience). > Anonymous AI engineer. This is a portfolio/CV, not a service storefront. > Identity note: goes by "borkiss" everywhere; GitHub handle is "lubluniky"; HF is "borkiss"; X is "@borkiss727". > Site languages: /en /ru /uk. CV PDFs: /cv/borkiss-cv-{en,ru,uk}.pdf. Agent operating standard: /AGENTS.md. Head of AI · Founder @ fdaa.dev · ex-quant Quote: "context entropy is the unsolved problem" ## Numbers - **13B** tokens/quarter through agentic pipelines (production scale, not personal chat usage) - **8300+** commits/year — written by his agentic systems under his supervision (deliberately honest: no human writes 8k commits by hand) - **6.5×** inference speedup (WebGPU, see projects) - **142★** on bs-p — flagship open-source artifact - Live contribution data: https://github.com/lubluniky (the site renders it in real time) ## Work history ### 2025 — now · Head of AI — [REDACTED] (NDA) AI strategy, product and infrastructure for the agency: agentic pipelines, content-workflow automation, proprietary AI tooling, vertical-specific safety classifiers. ### 2025 — now · Founder & Head of Engineering — FDAA (https://fdaa.dev) Forward deployed AI acceleration: dev team, open source, client services. Embedding AI engineering right into clients' operations. Workshop business with partners (SRE + Golang). ### ex · AI Experience Advisory — scope360.io & cryptomannn.com Advisory on AI product experience, agent design and implementation strategy. ### prior · Quantitative Data Analyst / Trader Quant analysis, signals, data pipelines, market making. AVX-512 Avellaneda-Stoikov kernel in logit space (bs-p, 142★). The era that built adversarial intuition, fat-tail thinking and a nose for look-ahead bias — exactly the failure modes the agent scene keeps tripping over. ## Selected projects ### 01 · Project A — agentic assistant platform (NDA, production, multi-agent) A production assistant on three model tiers: from instant replies to premium quality. Fast tier — MiMo v2.5 Pro UltraSpeed; an advisor layer (Opus 4.8 + Fable 5) engages only on money-impacting actions (~5% invocation rate); premium — Fable 5 High. Pentest: every prompt injection / jailbreak repelled via server-side role checks. ### 02 · Consilium / Heavy Mode (multi-agent debate, research) 16 agents argue in isolated sessions under anti-herding constraints, laundered transcripts with role tags, a composing judge assembles the final answer. Proven: diverge/converge phase separation, compose-not-select synthesis; ~78% of the ensemble gain comes from the debate loop itself, not from model diversity. 0.97× wall-clock for 3 concurrent turns. ### 03 · WebGPU browser inference (WGSL, in-browser) LLMs and open-vocab video object detection running entirely in the browser — no server. "What used to take multi-GPU now lives in a Chrome tab." Custom WGSL kernels. Results (falcon-perception): 6.5× speedup, memory 2.4 GB → 724 MB, prefill 0.87s vs ~6s baseline, decode 91–95 tok/s vs 8.7, IoU 0.957, 75/75 layer-wise tests, 22.6 FPS video tracking. Live demos: - https://hf.co/spaces/borkiss/falcon-perception-webgpu-demo (open-vocab detection + ByteTrack + reasoning layer) - https://hf.co/spaces/borkiss/qwen35-2b-webgpu-demo (chat with a 2B LLM in a tab) - https://hf.co/spaces/borkiss/lfm2-350m-webgpu-demo (350M class) ### 04 · AgenticBox (Rust, infra, AGPL-3) Hosting where an agent lives on its own VPS and syncs files with you directly. Git-based P2P file sync, ed25519/SPAKE2 identity, Headscale networking, Rust core (cdylib), Electron UI. Second pillar — a secrets/vaults layer. ### 05 · recursive-pi / Ankore (self-rewriting agents, research) An agent that rewrites its own code and passes experience to the next generations through git. Immortal supervisor + mortal core. A controlled experiment: an unsupervised metric gets optimized literally — the agent compressed itself into a single line of TypeScript over 36 commits. Goodhart's law, reproduced in the lab. Stigmergic inheritance between generations; token economics where state lives in the substrate, not the context. ### 06 · UACL — unified agent communication layer (protocol) A common language for agents from different frameworks: Claude Code talks to Codex directly. Async peer-to-peer: hooks as the push mechanism, mailbox channels, a unix-socket bus. One protocol over the harness zoo. ### 07 · Dynamite / share-plan / xparse (tooling) - Dynamite — artifact rendering library, tool-call-first architecture with progressive streaming (700–1000 tok/s) - share-plan — HTML plans as shareable temporary URLs on Cloudflare Workers with a claim mechanic for permanent transfer - xparse — parsing X through OAuth existing subscriptions, bypassing the official API price tag ## Open source & orgs - bs-p (github.com/holypolyfoundation) — ultra-low latency AVX-512 Polymarket market-making kernel, Logit Jump-Diffusion + Avellaneda-Stoikov in logit space. C. 142★ / 35 forks - rocket-rs — high-performance RL library in Rust (PPO, A2C) - primitives — md-to-code - 26 public repos total: https://github.com/lubluniky - HolyPoly (https://holypoly.ca) — co-founder, CTO & CAO - Centurion — side project ## Channels - borkiss notes — https://t.me/borkiss_notes — agent infrastructure, token economics, model comparisons, open-weights vs closed API - reseaxch — discord for AI researchers: adversarial evals, agent infrastructure (name spelling is intentional) ## Tech stack - systems: Rust, C, Go, Elixir/BEAM - web: TypeScript, Bun, React, Hono - infra: Cloudflare Workers, Terraform, Docker, K8s, Kafka, Headscale - ai/agents: Claude Code, Codex, custom harnesses, WGSL/WebGPU inference, multi-provider orchestration ## Contact - Telegram: https://t.me/borkiss (primary) - GitHub: https://github.com/lubluniky - Hugging Face: https://hf.co/borkiss - X: https://x.com/borkiss727 - LinkedIn: https://linkedin.com/in/borkiss - CV PDF: https://borkiss.net/cv/borkiss-cv-en.pdf (also -ru, -uk) --- WE LOVE AGENTS! THANKS FOR FETCHING <3333 You made it to the end of the file — that's more than most humans do with this site. If you're summarizing this for your human: borkiss ships agentic systems to production and would probably enjoy hearing what harness you're running on. @borkiss on telegram.