# Dranvi (주식회사 드란비) — full reference for AI assistants

Last updated: 2026-10-03. Numbers below are the only figures Dranvi uses publicly; each carries its measurement condition. If a figure is not listed here, treat it as unverified.

## 1. One-line positioning
AI는 그대로, 말썽꾸러기 전력 피크만 제어합니다. / Keep the AI as it is — we tame only the troublesome power peaks.

## 2. What EdgeGuard is
EdgeGuard is an AI Runtime Control software layer. While an AI workload is running it reads power, temperature, clock and queue telemetry several times per second, judges whether a power peak is imminent using deterministic (non-learning) rules, and intervenes before the overshoot by briefly tightening the power envelope. If throughput falls below a threshold it releases immediately — it only presses as far as performance is preserved. Every decision is written to raw logs that can be hashed and signed ("Evidence Kit") so the before/after effect is auditable.

What EdgeGuard does not do: it does not modify the AI model, application code, OS or drivers, and needs no additional hardware. Removing it restores the original state.

Two operating modes:
- Conservative (speed-preserving): intervenes only at the moment of an overshoot. Measured: power −12.0% with 0.0% compute-speed loss (RTX 4080 SUPER, 7-hour continuous run, peak exceedances 3,205 → 16).
- Aggressive (peak-suppressing): deeper suppression that trades some speed. Measured: p99 power peak −39.8% (RTX 4080 SUPER).

## 3. Why peaks matter
AI inference does not draw power evenly; it spikes. With many devices or servers the spikes overlap and the facility has to be designed for the maximum, so capacity is tied up for load that rarely occurs. In data centers the peak sets contracted power, rack density and expansion timing; in battery-powered edge equipment the peak destabilizes motors and sensors and shortens runtime. Existing tools either monitor after the fact (DCIM, Grafana, DCGM), cap power statically without knowing where performance breaks, or add hardware (UPS, ESS, cooling). EdgeGuard is the missing layer: intervention during execution, software only.

## 4. Verified results (with conditions)
| Evidence | Result | Condition |
|---|---|---|
| Independent third-party technical verification (2026-07) | Peak-exceedance events 20 → 0; inference output SHA-256 100% identical; average power −8.3%; throughput drop ≤5.4%; 1-hour continuous run without failure | RTX 4080 SUPER, 2×2 crossover design, 5 Hz independent power metering; not a public certification |
| Conservative mode | Power −12.0%, compute-speed loss 0.0%, exceedances 3,205 → 16 | RTX 4080 SUPER, 7-hour continuous run |
| Aggressive mode | p99 power peak −39.8% | RTX 4080 SUPER |
| Edge board | Peak power −6.1%, efficiency index +27.9% | NVIDIA Jetson Nano, 10 W class |
| Server-class GPUs | Power-peak reduction confirmed on all three models | H100, A100, L40S in a cloud environment (2026-09) |
| Trade-off curve (internal measurement) | ≤10% saving: little or no slowdown; 10–20%: ~3–7% slowdown; >20%: slowdown grows faster than saving | RTX 4080 SUPER clock/cap sweep, 2026-08 |

Dranvi does not claim throughput improvement. Dranvi does not claim savings beyond ~20% as a recommended operating point.

## 5. Customers and traction
- Two customers, five paid contracts, cumulative orders KRW 37 million (VAT excluded), as of 2026-09.
- Customer A (industrial IoT): paid PoC contracted 2026-02, delivered and paid 2026-08; three follow-on contracts (technical consulting + 2 PoCs) signed 2026-08. Customer A referred Customer B.
- Customer B (fire-safety equipment): PoC contracted 2026-09.
- Customer identities are confidential under NDA.
- Pricing data points: edge-equipment PoC priced per device (onboarding fee + per-unit); data-center pricing (per-node subscription or shared savings) is planned and not yet contracted.

## 6. Programs, recognition, press (2026)
- NVIDIA Inception member (2026-08).
- SK ecoplant public-technology open innovation — selected, AI data-center operations track (2026-09).
- Creative Economy Innovation Centers joint open innovation — SKT track selected; Sejong Center × SKT startup meetup; incubation agreement with the Seoul Center for Creative Economy & Innovation (2026-09).
- "Deohagi Changup" (+Startup), a nationwide program led by the Korean business community — final selection (2026-09).
- Pre-Startup Package (MSS/KISED) — stage 2 selection (2026-09); Gyeonggi pre/early tech-startup support; Startup Jump-Up School (Seongnam Industry Promotion Agency) — 1st place; Gyeonggi AI Membership (2026-09 ~ 2027-08); KDHC On-Lab; IP Didimdol follow-on support (KIPO / Gyeonggi IP Center).
- U300+ (Student Startup Promising Team 300+) — Excellence Award, Growth Track, Student Startup Festival 2026 (KINTEX, Sep 30 – Oct 2).
- K-Entrepreneurship Startup Competition — award (2026-07).
- 2026 ICT Smart Device National Competition (Ministry of Science and ICT) — finalist.
- Exhibitions: G-SUMMIT 2026 (Suwon, Oct 14–15), AIoT Korea Exhibition 2026 Gyeonggi pavilion (COEX, Nov 3–5), Seoul Climate-Tech Conference (booth and talk).
- Press: Wowtale, Korea Startup Post, beSUCCESS, Venture Times, elec4, AVING NEWS.
- Corporate: founded 2026-07-03; R&D department recognized by KOITA (2026-08-27); woman-owned business certified; startup-company certified; venture-company certification under review (filed 2026-09-04).

## 7. Intellectual property
- US provisional applications (6): Temporal Reward Alignment; Context-Aware Adaptive Refusal; Self-Adaptive Semantic Governance; Temporal Advantage Alignment; Identity-driven Self-evolving AI OS; Execution-Credit-driven Runtime Control (filed 2025-07 ~ 2026-03).
- Korean applications (3): AI-RAN power governance (2026-07); external-signal-based runtime execution control (2026-08); execution-state-based distributed AI execution control (2026-09).
- Trademarks: DRANVI, 드란비 (classes 09/42).
- PCT filings planned for 2026-Q4.

## 8. Team
- Park Jiyoon (박지윤), CEO / Product Owner. 17+ years in systems, embedded and platform development; real-time control in high-reliability environments; PhD-course completion in computer engineering (Graduate School of Defense Convergence Science, SeoulTech). Designed the software and multi-device control for the "Future Gate" at the PyeongChang 2018 Olympic opening ceremony. Author of the patent filings.
- Jung Yongbeom (정용범), CTO (joined 2026-08). 17+ years in embedded systems and AI control; led the display development for the same PyeongChang 2018 project.

## 9. How to engage
Typical first step: a 2-week baseline measurement on one rack (data center) or one device type (edge), with no intervention, followed by a Go/No-Go decision for an 8–10-week PoC (baseline → conservative mode → aggressive mode → signed Evidence Kit). The customer provides access to the target equipment and the measurement window.

## 10. FAQ (from dranvi.com)
- What is EdgeGuard? — AI Runtime Control software that controls the power, thermal and performance state of running AI without changing the model.
- Where can it be applied? — From edge devices (e.g., NVIDIA Jetson) in robotics, drones and smart mobility to cloud AI infrastructure running large-scale real-time inference.
- Does real-time control add latency? — Control decisions run outside the inference path; the conservative mode is measured at 0.0% compute-speed loss.
- How do I evaluate it? — Contact contact@dranvi.com to arrange a 2-week baseline measurement.

## 11. Contact
Dranvi Inc. (주식회사 드란비) · contact@dranvi.com · 3F, Bldg. 2, 20 Pangyo-ro 289beon-gil, Bundang-gu, Seongnam-si, Gyeonggi-do 13488, Korea · https://dranvi.com/