18. Building plasma hardware, autonomous research engines, and verification systems for model-generated work.
18. Currently building R0 Systems, plasma sanitization equipment for post-harvest agricultural produce, with a $30K validation deployment sold and commercial packhouse pilots targeted 2026. Prev. founded PlasmaX ($2M seed; decentralised on-site nitrogen production for ammonia synthesis), with $180K first production unit sold to CSIR-CFTRI India, 13 contracted deployments, 2 provisional patents filed, and 3 peer-reviewed publications. Also building Parchment Labs (autonomous literature-to-experiment pipeline) and delphi (pre-registered verification for LLM-generated quant strategies). Lavin Entrepreneurship Fellow at UW, YC AI Startup School 2026 SF cohort.
Founder & CEO . 2026 - Present
Building plasma sanitization equipment for post-harvest agricultural produce. Iterating on plasma reactor geometry, gas-flow control, and food-contact electrode design. Sold a $30K validation deployment; first commercial reactor unit and packhouse pilots targeted 2026.
Project Lead . Dec 2025 - Present
Built a 9-stage autonomous research pipeline: multi-source ingestion from ArXiv, bioRxiv, and Semantic Scholar across 50+ categories and 5,000+ papers/day; vector embedding via a local 32B-parameter model with HNSW indexing; novelty and relevance scoring with cosine deduplication; knowledge-graph gap detection; multi-agent adversarial peer review with 150 agents, weighted endorsement voting, and an 85% consensus threshold; experiment design; GPU execution; and LaTeX manuscript generation.
Project Lead . Jan 2026 - Present
Built a multi-stage verification system for LLM-generated quantitative strategies using a medallion-architecture data plane and a pre-registered promotion pipeline. Candidate strategies move through statistical screening, out-of-sample validation, and paper-trading stages, each with explicit acceptance criteria. Every rejection records the failed check, threshold, and realized statistic, creating an auditable dataset of LLM-generated artifact failure modes.
Lead Author . Mar 2026 - Present
Designed a timestamped, git-pinned protocol before any real-model run. The study benchmarks Qwen3-32B-Instruct and Llama-3.3-70B-Instruct across ForecastBench, LiveCodeBench v6, and GPQA Diamond with an 8-method optimization ladder including random search, coordinate descent, Bayesian optimization, successive halving, and genetic variants. The analysis plan uses linear mixed-effects models with equivalence bounds and Sobol-based sensitivity decomposition.
Lead Author . Feb 2026 - Apr 2026
Ran a reproducible study on reward-driven discovery failure in deep RL on procedural mazes. DQN, DRQN, PPO, and A2C with matched hyperparameter sweeps perform near a uniform random walk on 9x9 mazes, while a behavior-cloning MLP with identical architecture, observation, and optimizer reaches 97.4% test success. Initializing DQN from the BC-distilled weights and fine-tuning with the same shaped reward collapses test success from 97.2% to 13.6% across all 5 seeds.
Research Associate . Mar 2025 - Nov 2025
Worked under Prof. Alex Mamishev in the Sensors, Energy, and Automation Laboratory. Built a PPG drowsiness-detection wearable with a transimpedance front-end, 4th-order Butterworth filter from 0.5-5 Hz, 12-bit ADC at 100 Hz, and adaptive noise cancellation. Built piezoelectric hull-integrity sensing for a U.S. Navy program using PZT arrays, charge amplification, real-time 512-point FFT at 10 kHz, and acoustic-emission classification.
Research Intern . Jun 2024 - Mar 2025
Fabricated and characterized dye-sensitized and perovskite solar cells at the Centre for Sustainable Energy Technologies. Work included N719 on mesoporous TiO2, iodide/triiodide electrolyte, spiro-OMeTAD over MAPbI3, IV curves with Dyenamo Toolbox, photo-induced absorption spectroscopy for recombination kinetics, IPCE for spectral response, and XRD with Scherrer grain-size analysis.
Founder & CTO .
Engineered an industrial plasma system for decentralised on-site nitrogen production used in ammonia synthesis. Designed a custom 18 kV AC flyback driver at 23 kHz with under 2 W quiescent draw, plus an STM32-based PID control loop for gas flow, voltage regulation, thermal management, and MOSFET-driven power stages with real-time ADC feedback. Shipped $180K first production unit, contracted 13 deployments, filed 2 provisional patents, co-authored 3 peer-reviewed publications.
Model Craft Challenge . Apr 2026
Trained a 3.7M-parameter decoder-only transformer with 6 layers, 384 hidden dimension, and 6 heads on 12M FineWeb-Edu documents with a custom BPE tokenizer pipeline. Ran a 31-candidate factorial sweep on 8x H100 SXM varying tokenizer, architecture, optimizer, learning-rate schedule, warmup, weight decay, batch size, gradient accumulation, and data ordering. Divergent runs were watchdog-terminated and pairwise config selection used Welch's t on BPB with wallclock-aware tiebreaking.
Creator . Feb 2026
Open-source multi-agent LLM runtime with 4,400+ LOC across Python and TypeScript. It supports declarative YAML agents, LiteLLM routing across 8 providers, pluggable memory with SQLite and ChromaDB, 6 built-in tools, sequential/parallel/supervisor patterns, and FastAPI plus WebSocket observability with a Next.js dashboard.
Creator . Apr 2026
Open-source continuous-learning framework across Rust, TypeScript, and Python implementing biologically inspired neural regions with local Hebbian learning rules and neuromodulatory reward signals. It uses no backprop, optimizer, or training loop. The system includes prioritized experience replay with TD-error-weighted sampling, associative and episodic memory with persistent cross-session state, and a safety kernel that gates dangerous actions before execution.
Creator . Apr 2026
Teachable Machine re-imagined for modern foundation models. WireML runs browser-first via WebGPU with zero-backend hosted mode, while the Python runtime auto-detects CUDA, MPS, MLX, ROCm, DirectML, XPU, or CPU. Data sources, backbones, training heads, evaluators, and deployers are draggable nodes. CLIP, DINOv2, SigLIP, Whisper, and MediaPipe ship out of the box with Hugging Face Hub support.
Creator . Apr 2026
Platform for converting saved content into compounding technical leverage. It ingests bookmarks, highlights, articles, and notes into a queryable knowledge graph with semantic search, enabling reasoning across disparate sources.
Creator . Mar 2026
Real-time multi-agent coordination in TypeScript: WebSocket relay with automatic peer discovery, shared context store with causal ordering, priority task queue with work-stealing, adversarial code-review protocol, sandboxed remote command execution, and cross-network auto-tunneling.
Creator . Mar 2026
Persistent multi-agent simulation with a real economy, trading, resource management, emergent social dynamics, and simulated urban interaction. Agents develop relationships, trade, form organizations, and evolve behavior over time.
Creator . Mar 2026
End-to-end robotics imitation-learning pipeline with 41 modules and 103 tests. It streams a MuJoCo Franka Panda 7-DOF simulation at 60 FPS via FastAPI and aiortc WebRTC, uses damped Jacobian pseudoinverse IK, records with LeRobot v3 to Parquet and MP4, trains Action Chunking Transformer and DDPM Diffusion Policy with cosine noise, exports ONNX, and integrates with Hugging Face Hub.
Creator . Mar 2026
Real-time accelerometer-to-bellows mapping with dynamic filter sweep, musette tuning via three detuned oscillators per note, two octaves of chromatic treble keys, six bass/chord buttons, and a self-signed HTTPS local server for iOS accelerometer access.
Creator . Feb 2026
100% local relationship analytics engine. No data leaves the device. On-device ML analyzes communication patterns, attachment styles, and relationship dynamics from chat screenshots while preserving privacy.
Project Lead . Nov 2025 - Present
Built open-source underwater robots for lake-health monitoring: vectored 5-thruster configuration on Pixhawk 4 and ArduSub with MAVLink, RPi 5 H.264 video via GStreamer over CAT6 tether under 100 ms latency, custom KiCad MOSFET power distribution with 4x IRF3205, 6S Li-ion 22.2 V / 5 Ah battery, current sensing, over-current protection, and waterproofed acrylic housings rated to 100 m with dual O-ring seals.
Creator . Apr 2026
Personal portfolio site for showing the systems, research, and stories that a traditional resume falls too short for. Built with Next.js, TypeScript, structured content data, and an editorial profile surface.
Engineer / Researcher / Founder
Currently building R0 Systems, plasma sanitization equipment for post-harvest agricultural produce. Prev. founded PlasmaX ($2M seed; decentralised on-site nitrogen production for ammonia synthesis). Also building Parchment Labs (autonomous literature-to-experiment pipeline) and delphi (pre-registered verification for LLM-generated quant strategies).
Now R0 Systems / post-harvest plasma sanitization
also Parchment Labs + delphi
Education
2024 to 2028
University of Washington, Seattle
B.S. Electrical & Computer Engineering + Applied Mathematics
GPA 3.93/4.0 · Lavin Fellow · YC AI Startup School 2026
Highlights
BUILDING
R0 Systems
Plasma sanitization equipment for post-harvest agricultural produce. $30K validation deployment sold; first commercial reactor unit and packhouse pilots targeted 2026.
PREV. SHIPPED
PlasmaX
$2M seed at $8M post-money. Decentralised on-site nitrogen production for ammonia synthesis. $180K first production unit sold to CSIR-CFTRI India, 13 deployments contracted, 3 peer-reviewed publications, 2 provisional patents filed.
RESEARCH
Parchment Labs
Autonomous literature-to-experiment engine. Local 32B inference, HNSW retrieval, 150-agent adversarial peer review, GPU execution, LaTeX manuscript generation.
VERIFYING
delphi + agentbreed
Pre-registered verification loops for LLM-generated quant strategies and multi-agent LLM configuration studies. Structured rejection logs, test-gated evaluation harnesses.
LAB WORK
UW SEAL Lab · NIIST-CSIR India · Maze RL
Embedded sensing for U.S. Navy hull integrity, PPG drowsiness wearables, solar-cell characterization, reproducible RL baselines showing reward-driven discovery failure.
Selected Recognition
YC AI Startup School 2026 · San Francisco cohort
OpenAI Parameter Golf · beat prior 2nd-place track_10min_16mb
Google Scholar · 35 citations, h-index 3, i10-index 2
Carnegie Mellon Venture Challenge · 2nd / 150+ teams · $4,500
UW Science & Technology Showcase · Grand Prize + Best Pitch · $2,750
Red Bull Basement · Runner-up
Lavin Entrepreneurship Fellow · four-year selective program
Todo (Not That Kind)
