kula — AI Systems Engineer

60s → 263msagent hot path, ~150× faster, A/B-verified

2nd placenational copyright org's AI-music-detection procurement

3 days → 5 minAI video production pipeline, sole engineer

5M+ recordskeyword search migrated to pgvector semantic search

P1 criticalBinance bug bounty — accepted report (Bugcrowd)

I build AI systems that survive contact with production. Twenty years shipping software; the last three spent entirely on the part most engineers skip — latency budgets, provider failover, evaluation harnesses, and the infrastructure that keeps an agent answering in under 10 seconds instead of 39. I read the codebase before I propose the plan: my last provider migration shipped as two environment variables per service instead of a multi-repo rewrite, because an audit showed all three AI surfaces already shared one hook pattern.

Selected Work

AI Agent Platform — venture-backed marketing SaaS (contract, ongoing)

Cut a 41-tool LLM agent's curation hot path from a 60s timeout to 263ms (~150×), quality verified with an A/B evaluation harness on production workloads.
Migrated 5M-record creator search from ILIKE keywords to pgvector semantic search; stood up a prod-parity staging gate with deterministic E2E fixtures.
Designed a central LLM gateway migration (100+ providers) that shipped with zero application code changes — env-only, per-service.
Surfaced and reported 4 production security issues during the audit (plaintext GitHub PAT in a deploy URL, publicly exposed gateway port, invalid failover key, shared master key).
Unified two divergent agent routes (3,262 + 1,297 lines) onto one shared harness, sequenced by per-function risk analysis.

Next.js · DeepSeek/Gemini · pgvector · PostgreSQL · LiteLLM · Docker

AI-Generated-Music Detection — national copyright organization bid (2nd place, procurement evaluation)

3-detector consensus engine: spectral fakeprint (ISMIR 2025) + SpecTTTra spectrogram transformer (ICLR 2025) + MERT v1-95M embedding head.
3-band routing policy — auto-pass / auto-flag / human review on detector disagreement — instead of forcing verdicts.
AudioSeal watermark evidence layer that overrides model consensus; Demucs 4-stem per-instrument scoring.
False-positive correction tuned for Korean/Chinese vocals — the differentiator vs. the incumbent. 66 tests, Dockerized, VPS-deployed.

Python · PyTorch · ONNX · FastAPI · Demucs · Docker

Cenema — AI video automation platform (CTO / sole engineer)

Topic → scene planning (Gemini) → image gen → video gen (Kling) → NLE export (CapCut/Premiere). Production time ~3 days → under 5 minutes.
Orchestrated 12+ AI providers with health-check-driven failover: 99.9% availability, −30% operating cost.
Skia-canvas + FFmpeg render engine, 70× faster than the Puppeteer approach it replaced.

TypeScript · Cloudflare Workers · Inngest · Gemini · Kling

AI travel planning platform (contract + equity, ongoing)

Built the AI itinerary planner with a provider-selection layer (DeepSeek and multi-provider routing) and free-form stop editing.
Shipped payment refund safeguards and legal-disclosure compliance — the money path, engineered defensively.
Place-data pipeline + CDN image optimization.

Earlier (2004–2024)

Enterprise on-prem RAG document platform (CTO) — Qdrant, LlamaIndex, LangChain.
Futures auto-trading / HFT systems with Hyundai Securities; enterprise delivery for LG and SK.
Accepted P1 (critical) report, Binance bug bounty (details under disclosure limits).
Founder/PO of a mobile product grown to 1.5M+ downloads (2007–2018).

How I Work

Spec-first: every change starts as a written proposal with the problem, the measured baseline, and what I'm explicitly not touching. Never claim an optimization without an eval behind it.