Two problems that look unrelated but share one shape: the model is rarely the bottleneck — the data, and the discipline around it, is what decides whether anything holds up.

Research and trading infrastructure for prediction markets and weather / temperature markets.

Market data. Tick-level collection and normalization across linked markets; order book, trade and fill reconstruction. Resolution rules and fee schedules are captured next to prices, so a backtest cannot silently cheat on settlement or costs.

Strategy research. Executable-arbitrage scanning across correlated contracts. Regime-transition analysis to see where a strategy's assumptions break. Inventory-signature and fee/rebate pairing to separate genuine edge from subsidy. Residual-PnL attribution to surface what the model is not explaining — usually the part that matters.

Validation. Replay-based fitting (spec versus strategy), held-out behavioral evaluation, and live-cycle evidence collection. A strategy has to survive data it has never seen before it is allowed size.

Execution. Latency-aware order placement, taker-trigger probing and coverage sweeps, continuous position monitoring, and explicit risk gates.

Data is the bottleneck in robot learning, not the architecture. I work across the pipeline that feeds it.

Collection. Teleoperation, including browser-based and crowdsourced setups; handheld and mobile IMU capture; human–object interaction reconstruction.

Video reconstruction. Recovering scene geometry, camera motion and object interaction from video, then rebuilding it as simulation-ready assets and trajectories.

Alignment. Multimodal spatiotemporal alignment, camera intrinsics, visual–inertial odometry and SLAM — making heterogeneous streams agree on where and when.

Simulation. MuJoCo and Isaac Lab are the two simulators I work in — scene and asset construction, contact and actuation modelling, and synthetic episode generation.

Representation. Episode and trajectory design, 3D spatial representations, cross-embodiment latent actions, standardized episode formats.

Evaluation. VLA model evaluation, data-quality rubrics, sim-to-real transfer, and world-model planning evaluation.

The systems underneath both directions, built end to end in TypeScript and Python:

• Real-time data systems — high-frequency ingestion, streaming processing, and reproducible experiment harnesses.
• Data platforms & pipelines — collection → processing → export; multi-region streaming workers and asynchronous job queues.
• Telegram ecosystem & Mini Apps — multi-bot clusters on shared runtimes, wallet and payment flows, admin consoles, i18n.
• Device automation — real-device RPA and mobile data acquisition.