ml infrastructure
tensorforge-webgpu
9 tests; the default graph lowers from 13 to 8 nodes and the live GPU result is checked against a seeded CPU oracle.
8 → 3primitive kernels to GPU dispatches
VERIFIED 2026-08-09 RECRUITER ENTRY / 90 SEC
I build the layer
underneath the model.
Systems and ML infrastructure engineered with explicit failure modes, reproducible commands, and the benchmark boundary published beside the result.
- Measured systems
- 24
- Test functions
- 815
- Live labs
- 5
ARCHITECTURE SIGNAL ROUTE ACTIVE
MLI / ML InfrastructureCan the model-serving layer earn its speedup?Retrieval · scheduling · tensor compilation
01TENSOR02GRAPH03MEMORY04GPU05PROOF
ACTIVE PROOF PATH / MLI
Three systems. No vague claims.
Select a route above. Every result links to a proof passport, runnable source, published limitation, and—where available—a live browser lab.
ml infrastructure
annlite
51 tests; at 98.5% recall it performs 26x fewer distance computations than exact scan.
1.83xFAISS QPS at 0.999 recall
ml infrastructure
vllm-lite
90 tests; paged scheduling reaches 2.9x sequential throughput and 5.5x better median TTFT.
94%KV slot utilization vs 21%
FRONTIER BUILD / LIVE WEBGPU
TensorForge turns compiler internals into something recruiters can operate.
Typed tensor IR, shape inference, identity canonicalization, operator fusion, output-rooted dead-code elimination, liveness-aware buffer reuse, and three generated WebGPU compute kernels.
13 → 8graph nodes
8 → 3primitive kernels to GPU dispatches
CPUnumerical oracle
ENGINEERING SURFACE
One verification spine across four domains.
Storage + consensus
MVCC, Raft, vectorized execution, deterministic fault injection.
ML infrastructure
HNSW retrieval, paged KV scheduling, distributed training, GPU compilation.
Autonomy systems
Signal processing, radar image formation, association, VIO evaluation.
Evidence engineering
Independent oracles, paired seeds, invariants, explicit failure boundaries.
CHANNEL OPEN