Adam Ctverak

HyperWatch, San Francisco. Co-founder. Led electro-mechanical design and flight-test qualification.

Redundant flight-termination subsystem on HyperWatch’s high-altitude balloon platform. The platform provides IR search-and-track for hypersonic missile defense.

Problem: FAA Part 101 §101.35 required two independent payload cut-down devices. No COTS unit fit the mass, reliability, or integration envelope.

Built: 3D-printed burn-wire mechanism. Nichrome element severs a braided Spectra (UHMWPE) tether on command. 2S LiPo direct-drive, sized from a CV/CC buck characterization rig. Active IMU-based de-spin eliminates wire winding. Concentric geometry, minimal wire slack. Custom pull-test fixture.

Pull-tested to 13.25 kg, >2× max flight load. Four flight campaigns. Zero failures.

Company: $80K DIU award, $250,000 from venture capital, Founders Inc Blueprint (Fall 2025).

Tool: hypersonic-trajectory-simulator, balloon-vs-ground sensor trade studies on representative boost-glide trajectories.
Origin: Stanford H4D, DIU Summer Fellowship 2024.

STMicroelectronics STEVAL-EDUKIT01. Personal project, in progress alongside Stanford AA203 (Optimal & Learning-Based Control) and AA212 (Nonlinear Control Design).

Assembled the Furuta pendulum to chain advanced control end-to-end from theory to hardware. Curriculum: PID → LQR → MPC → energy-based swing-up, each gated through a custom Python sim and the physical board.

Problem: Most controls coursework stops at simulation. Real hardware exposes the gap between idealized plants and what actually runs: motor saturation, encoder quantization, sample-period latency, friction, and the coupling that the linearized model drops.

Built: STM32F401RE Nucleo board flashed with stock STSW-EDUKIT01 firmware. Captured a hardware log under open-loop rotor sweep with the pendulum hanging passive, sampled at 200 Hz. Implemented Module 1 (Classical) in Python: PD design on the rotor double-integrator plant G(s) = 1/s² from Jθ̈ = τ via Newton’s second law, closed-loop second-order analysis, and overshoot/settling specs. Plot script auto-selects the most active 12-second window from the log and overlays the rotor command, the measured rotor angle, and the passive pendulum’s damped response.

Module 1 sim and hardware baseline complete. Modules 2–4 (LQR upright, linear MPC, energy-based swing-up with LQR catch) on the roadmap, derived Newton-Euler from each link.

Code: github.com/nadranos/rotary-pendulum-control.

Astroport Space Technologies, San Antonio, TX. NASA STTR. Lab intern. Led the temperature-strength correlation study on CSM-LHT-1 lunar simulant.

Problem: Autonomous lunar landing pad construction needs an in-situ material that beats concrete on strength. No published map between thermal-cycle parameters and compressive strength for CSM-LHT-1.

Built: Induction furnace melts on 75 g CSM-LHT-1 charges in graphite crucibles. Test matrix spanned 8–33°C/min ramps, 1150–1310°C peaks, quench vs 10°C/min anneal. Custom retention fixtures (modified ASTM C1231) with neoprene pads to align cores. Uniaxial compression per ASTM C39 on a 100 kN MTS at 0.014 mm/s. Helium pycnometer density at UTSA HAMsTER lab.

43 MPa mean compressive strength, 98 MPa peak. Bricks ran 54% stronger than NASA Launch Pad 39B concrete. Closed porosity 0.68–2.55%. Primary failure mode: longitudinal splitting. First-author poster at Texas Area Planetary Science (TAPS) 2023.

Poster: TAPS 2023 [pdf].
Coverage: LPI / USRA feature on TAPS 2023.

Stanford CS129 (Applied Machine Learning). Solo.

Component-degradation prediction for small modular reactors, trained on a Pebble-Bed High-Temperature Gas-Cooled Reactor (PB-HTGR) Simulink digital twin.

Problem: SMRs depend on early degradation detection. Threshold-based anomaly monitoring misses gradual failures. Can a supervised-learning pipeline predict per-component degradation (0–5%) from multivariate sensor streams?

Built: Data pipeline on a PB-HTGR Simulink digital twin (Rivas et al., NC State). 65 simulation scenarios across steady-state, load-following, and up/down-ramp modes, Gaussian-noised sensors, >50 features, 352K tabular samples. Five components labelled 0–5% (Circulator Pump, Feedwater Pump, Condenser Pump, HPT, LPT). Three models: XGBoost on tabular data, bidirectional LSTM with attention and per-component output heads, and a confidence-weighted hybrid with dynamic per-label weighting. Engineered features: physics-based component efficiencies, time-derivative response.

XGBoost: 75.1% overall accuracy. Hybrid: 73.3%, beat XGBoost on HPT, FWP, and LPT individually. Bidirectional-attention LSTM underperformed at 14.6%, bounded by sequence-sample scarcity (140 sequences vs 352K rows). All models weaker on intermediate degradation states than 0% and 5% extremes.

Code: github.com/nadranos/smr-predictive-diagnostics.

Florida Institute of Technology, Senior Design (AEE 4292). Project Manager, 9-person team. Owned schedule, budget, and test-plan review.

Autonomously deployable hybrid inflatable lunar habitat, sized to fit the SLS Block 2 cargo payload envelope. Full-cycle capstone from Statement of Work through scaled engineering-model demonstration.

Problem: Artemis-class missions need habitats that ship volume-efficiently on SLS and deploy autonomously on the lunar surface. Rigid habitats waste payload volume; collapsible structures buy more surface infrastructure per flight.

Built: 8-arm radial structure in 30x30 T-slot aluminum extrusions with double-shear butterfly joints, PVC/stainless center column, 8 folding baseplate pairs. Primary deployment: motor-driven worm drive in the center column pulls tension cables through the upper T-bar track to lower the arms. Secondary deployment: spring-loaded T-pins release the baseplates on ground contact. Arduino-based ECLSS with BME280 sensing and solenoid-driven pressurization. Full Ansys static structural FEA on arms, joints, and baseplates. 2.2:1 expanded-to-collapsed volume ratio.

PDR and CDR passed. Scaled engineering model fabricated and fully deployed on camera at the Spring 2024 Florida Tech Senior Design Showcase, $2,000 total budget. Design pitched to and evaluated by Blue Origin engineers.

Personal project. Solo build. Sole operator.

Two-division intraday trading system on the open-source OpenClaw multi-agent gateway: a research loop that proposes and scores strategies, and a Python daemon that executes the survivors against a paper Interactive Brokers account.

Problem: Intraday strategy research is bottlenecked by hand-coded backtests and ad-hoc evaluation. Live execution is bottlenecked by the gap between backtest assumptions and broker reality (slippage, partial fills, Greeks staleness, regime drift).

Built: Experiment registry in SQLite (RomulusDB, 1,376 lines) tracking 458 experiments across 82 strategy families and 3 asset classes. Composite reward function R, z-score-normalized across the population over 11 metrics (per-trade Sharpe, profit factor, max intraday drawdown, holding time, etc.) with explicit reward-hacking flags. Live execution daemon RIA (~5,500 lines) on ib_async, with regime-aware strategy selection and asset-class executors for equities (ORB+VWAP, momentum, gap fade), forex (triangular arb, MACD, mean reversion), and options (IV-surface mispricing, Kalman, gamma scalping with full delta/gamma/vega/theta from Massive.com). Risk manager with position-size, exposure, daily-loss, and trade-count gates plus a kill switch that flattens every position with market orders. Bar aggregator converting 5s ticks into rolling 1-minute OHLCV with cumulative VWAP. Backtest replay validator that proves live signal handlers reproduce backtest entries bar-for-bar. Operator dashboard (Python HTTP server, dark-terminal HTML) streaming summary stats, leaderboard, R-trend, and live execution state. Six-agent control plane on OpenClaw with 30-minute heartbeats, dispatched directives, and Discord telemetry.

Live paper trading on a $10K IBKR account since February 2026. 156 of 458 backtested strategies passed the gate and reached the live executor. End-to-end loop verified: backtest selection, live execution, performance feedback, and selector re-weighting via the live-vs-backtest ratio.

Code: github.com/nadranos/romulus-trading.