PyTorchReverse EngineeringPPO
Reinforcement Learning Agent with Memory Reverse Engineering.
Reverse engineered game memory layout via signature scanning to extract coordinate hashes and state bytes; bypassed anti-tamper protections to enable real-time RAM reads through a custom DMA python module using pymem.
Trained a DNN + decision-tree hybrid (minimax-style lookahead) via multi-agent PPO self-play, achieving superhuman play. Eliminated SubprocVecEnv IPC overhead with a vectorized MK11VecEnv for ~7.5x training throughput.
[ Currently Closed Source ]
PythonPyTorchMulti-GPU
A 4M-parameter LLM built entirely from scratch.
Trained an 8-head Transformer from scratch on a single RTX 5070 over 3 weeks. Achieved multi-sentence coherent dialogue via mixed-precision training and gradient checkpointing with no pretrained weights.
Optimized custom tokenization, batching, and data pipelines to improve inference quality and training throughput.
[ Currently Closed Source ]
PyTorchJetson NanoEdge AI
Local voice assistant running full keyword detection and intent parsing.
System runs on an NVIDIA Jetson Nano, combining a 5-million-parameter PyTorch conversational model with Picovoice’s Porcupine for keyword spotting and Rhino for intent parsing.
Executes natural language understanding fully on-device, mapping recognized intents to GPIO or MQTT-based appliance controls with zero cloud dependency.
View Source Code →
ESP32PID ControlC++
Hybrid temperature regulation system for ultra-stable thermal management.
Achieved ±0.05°C stability on an ESP32 using a dual-mode control system (lag-aware PID + constant-power model) with predictive braking and sensor-lag compensation.
Enabled rapid convergence and runtime tuning through a serial control interface with CSV and multi-sensor capabilities.
View Source Code →
C++ESP32TFLite Micro
Knowledge distillation for low-power edge inference.
Embedded DSCNN based ML system for emergency siren detection: distilled a 4M-parameter TensorFlow teacher into a 38k-parameter DSCNN on ESP32.
Achieved 99.97% test accuracy, 92% real-world accuracy, 18 ms latency, and 3+ days per 1% battery using a quantized, low-power inference pipeline with optimized DSP preprocessing.
[ Currently Closed Source ]
ReactNode.jsWebSocketsDHKE
Secure, encrypted full-stack collaboration network.
Built a 30k-line, 165-commit encrypted platform featuring DHKE chat, project posts, auto-groups, tasks, and threads.
Eliminated plaintext exposure via salted SHA-256 auth, per-session DHKE keys, and optimized WebSocket messaging.
[ Currently Closed Source ]
Applied CryptographyCurriculum DesignCTF
Designed a graduate-level cryptography dojo on pwn.college.
Authored 59 hands-on challenges across 10 modules spanning modular arithmetic, RSA, block-cipher exploitation, PRNG attacks, lattice-based cryptanalysis, ECC, and zero-knowledge proofs.
Used to teach a graduate cryptography course, with 3,600+ solves and 500+ unique hackers working through the material to date.
View Dojo →
Multi-Agent SystemsAutomationCTF
Agent-driven harness for automated CTF challenge solving.
Helped build an agentic solver harness used by Shellphish to automatically triage, reason about, and exploit CTF challenges (pwn, rev, crypto) with minimal human intervention.
Focused on orchestration logic connecting LLM reasoning agents to real exploitation tooling — debuggers, disassemblers, and exploit primitives — for end-to-end autosolving.
[ Internal — Shellphish Infra ]
Adversarial MLModel QuantizationSecurity
Training quantized surrogate models to craft transferable adversarial attacks.
Wrote the training and exploit scripts for a quantized extension of the Zero-Query Boundary Attack (ZQBA), building substitute models over CIFAR-10, CIFAR-100, and Tiny-ImageNet to study transferable black-box attacks.
Used the trained surrogate models to craft adversarial examples that transfer to real, full-precision target models — extending the codebase as a foundation for follow-up zero-knowledge-proof research.
View Source Code →
LLM ProductFull-StackApplied AI
An AI-native dating platform, now spinning out into a startup.
Designed the AI matching and conversation layer end-to-end, using LLM-driven profile understanding and compatibility signals in place of static swipe mechanics.
Currently transitioning from prototype to an incorporated startup.
[ In Development — Stealth Startup ]
