Encrypted ML inference engine (Rust + Python)

A privacy-preserving ML inference library that runs PyTorch / sklearn / XGBoost models under Fully Homomorphic Encryption — inputs and outputs never leave ciphertext. A 'narrow-waist' 3-layer architecture lowers ONNX graphs to a versioned IR executed on tfhe-rs primitives, with a PyO3 bridge and bit-for-bit exactness guarantees.

Distributed LLM inference engine in C++20

A distributed LLM inference engine that runs models like Llama 3 70B across a ring of heterogeneous consumer devices using pipeline parallelism — bypassing the VRAM wall. Hand-written AVX2/NEON SIMD kernels, a zero-copy Linux kernel-module transport, and a custom INT8/FP32 weight serializer.

Twin-fleet incident response agent with formal safety verification

An incident-response agent that rehearses candidate remediations across isolated Kubernetes twin environments receiving live-mirrored traffic before touching production. A deterministic rubric scores candidates in parallel, while a Z3-backed SMT safety kernel formally verifies system invariants to veto unsafe actions.

Intelligent LLM context-compression engine

A high-performance LLM input-compression framework. A fine-tuned BERT token classifier scores every token by semantic necessity, then a two-tier pruning engine (context + token level) cuts prompt size while a zero-hallucination constraint engine preserves code and structure.

Low-latency prediction-market logical arbitrage engine

A low-latency algorithmic trading system that detects and exploits mathematical inconsistencies across 1,000 live Polymarket order books in ~14ms. An LLM reasoning core (watsonx.ai + LangGraph) maps logical relationships across prediction markets, coupled with a Rust execution engine featuring Fill-or-Kill settlement, circuit breakers, and fixed-point math.

Local-first terminal-native AI job-application co-pilot

A local-first, terminal-native job-search platform featuring an interactive Textual TUI, Typer CLI, and background daemon over SQLite (WAL). A pluggable provider abstraction orchestrates Claude Code, Codex, and LiteLLM models with a 5-rung recovery ladder, truth-anchored resume tailoring, and a 100% line and branch test coverage gate.