01 / WORK
Built, broken, still running.
The source is there. So are the rough edges. Each plate maps the system that actually exists—not a metaphor for it.
01 / LOCAL_REPO / FIRST_TURN_VERIFIED STATE MACHINE / RUST AUTHORITY
ARES
Civilization Agent Sandbox
Agents submit bounded, typed proposals; Rust validates the base tick and full allocation, checks invariants, and owns the only authoritative state transition. The first settlement turn is deterministic and atomic.
- 01 Bounded proposal
Complete six-part allocation plus the expected base tick.
- 02 Rust validates
Build a candidate state, then check every invariant.
- 03 Commit or reject
S(0) becomes S(1) once; rejection leaves state untouched.
02 / PUBLIC_REPO / WORKING_BUILD CONTROL PLANE / LEASE RECOVERY
Rust Edge Compute Node
Distributed Compute Control Plane
I built a fault-tolerant control plane for distributed compute. Workers report over gRPC; the orchestrator handles leases, priorities, and recovery.
- 01 Workers report
GPU telemetry, uptime, version, and heartbeat state.
- 02 Orchestrator leases
Capability filter, priority queue, and renewable job lease.
- 03 Failure recovers
Stale nodes are evicted and expired work returns to the queue.
03 / PUBLIC_REPO / SPATIAL_HASH_TODO SOLVER CHAMBER / CURRENT CPU PATH
High-Performance Physics Engine
Particle Simulation
A C++17 2D particle engine with Verlet integration, boundary constraints, diagnostics, and naive O(n²) particle contacts. The spatial hash and honest benchmarks are still next.
- 01 Integrate
Current and previous positions advance through Verlet steps.
- 02 Resolve
Boundary bounces and all-pairs particle contacts on CPU.
- 03 Render + measure
SFML frame output with FPS and particle diagnostics.
04 / PUBLIC_REPO / VERIFIED_MVP EXPERIMENT BOUNDARY / ROUTE + MEASURE
Speculative Inference Proxy
Speculative Decoding Gateway
A functional Rust gateway that routes identical OpenAI-compatible requests across baseline and speculation-enabled vLLM, streams SSE without buffering, and measures whether speculation actually helps.
- 01 Route
One request selects the baseline or speculation-enabled arm.
- 02 Engine owns speculation
Draft proposal and target verification stay inside vLLM.
- 03 Stream + measure
Pass SSE through while recording first output and total latency.
02 / STACK
What I use to build.
Things I’ve used to build, debug, and measure real systems.
C_01
Systems programming
Control planes, runtimes, and hot loops.
- Rust
- C++
- Linux
- Bash
- Tokio
C_02
AI infrastructure
Model serving, evals, and finding out where the time went.
- Python
- PyTorch
- vLLM
- FastAPI
- Eval Harnesses
C_03
Distributed infrastructure
Schedulers, telemetry, recovery, and the failures between them.
- Docker
- gRPC
- Tonic
- GitHub Actions
C_04
Security labs
Threat models, adversarial inputs, and labs built to be broken.
- Hack The Box
- Web Security
- Prompt Injection
- Threat Modeling
C_05
Interfaces
Interfaces for seeing what a complicated system is actually doing.
- TypeScript
- React
- MapLibre
- Dashboards