Low-level systems · On-device AI acceleration · Decentralized networks

I build systems at the boundary of C-level performance engineering and applied ML: inference engines on Apple Silicon, single-file P2P distribution, and ML pipelines measured against reproducible baselines. The constraint is usually physical — memory bandwidth, cache residency, verification cost, or network interference.

github.com/ayush6288 · Raipur, India

C11 · ARM NEON · Apple Silicon · Source

Problem & architecture.

Autoregressive decoding on edge hardware is bounded by per-token weight streaming and the cost of verifying draft tokens. SPEL is a bare-metal C11 engine that pairs a 2-bit ternary draft with a 12-layer hybrid target — 9 Gated-DeltaNet recurrent layers and 3 RoPE attention layers — verifies K+1 tokens in one batched pass, and rolls back attention KV and recurrent state transactionally.

Engineering highlights.

• 1.45×–1.48× net decode speedup over autoregressive generation at speculative depth K=4, measured on the 21M-parameter trained checkpoint with 4 draft threads and an in-domain prompt.
• 2.7×–3.1× the FP32 per-MAC rate from hand-written, vectorized NEON 2-bit ternary GEMV kernels on a 4096×1536 shape. Packed weights occupy 1.50 MiB versus 24.00 MiB in FP32.
• ~0.5 ns commit/rollback for attention and recurrent state, against 108.2–112.7 µs to copy the 21M model's 4.50 MiB KV cache.
• Bit-identical verification in strict mode, with a dynamic speculation governor that detects low acceptance and falls back to autoregressive decode.
• Persistent draft thread pool removes per-GEMV thread creation; 4-worker draft throughput reaches 4,394–4,628 tok/s.

Stack / primitives. C11 · ARMv9-A NEON (vdotq_s32) · Apple Accelerate/AMX · pthreads · packed 2-bit weights · int8 activation quantization

Raw captures and derived rates are committed in results/.

TypeScript · WebRTC · P2P · Source · v0.3.0

Problem & architecture.

Centralized publishing fails at the origin, the CDN, or the account layer. SPORE is a local-first, censorship-resistant distribution client compiled into one self-contained HTML file: signed packets propagate over WebRTC DataChannels in room-scoped swarms, with no required server, database, or origin, targeting high-interference networks where public signaling may be blocked.

Engineering highlights.

• Single ~320 KB artifact with zero network requests at boot; runs from file:// and serves embedded media with no peers online.
• Local verification only: secp256k1 Schnorr signatures, SHA-256 packet IDs, and Hashcash proof of work. Unknown authors must clear a 16-bit work floor before entering the live feed.
• Chunked media transport: 64 KiB SHA-256-verified chunks over WebRTC DataChannels, 4–6 chunk request windows, and deterministic room scoping per creator and media hash.
• Offline payload paths: lossless RGB LSB steganographic PNG carriers and HTML quine snapshots carrying up to 200 verified packets.
• Interference tolerance: BitTorrent WebSocket trackers and Nostr relays for signaling, plus a compressed manual SDP bridge when public signaling is unavailable.
• Documented threat model: Dandelion++-style source obfuscation with explicit non-goals — not anonymity, not confidentiality.

Stack / primitives. TypeScript · WebRTC DataChannels · secp256k1/BIP340 Schnorr · Hashcash PoW · IndexedDB · PNG LSB steganography · Vite · Vitest

Python · PyTorch · MPS/Metal · Private repository

Problem & architecture.

ML prototypes usually break between the notebook and the service boundary: feature logic diverges, inference depends on a cloud round-trip, and evaluation is added after deployment. This pipeline treats ingestion, feature materialization, local inference, and evaluation as one system with shared schemas and explicit interfaces.

Engineering highlights.

• Vectorized feature store with one schema shared by training and serving, removing train/serve skew at the boundary.
• Local MPS/Metal inference path for on-device execution, keeping the hot path off the network and the runtime footprint bounded.
• Lightweight evaluation at every stage — ingestion, feature materialization, inference latency, and output quality — with regression checks before release.
• REST and gRPC interfaces with versioned request/response contracts and explicit service boundaries.
• Runtime footprint treated as a first-class metric alongside model quality: artifact size, startup time, and peak memory.

Measured surface. Ingestion throughput · feature materialization latency · inference p50/p95 · artifact size · peak RSS · evaluation delta per release

Stack / primitives. Python · PyTorch · MPS/Metal · FastAPI · gRPC · SQL · vectorized feature batches

• Pair every performance claim with a baseline and a committed capture.
• Optimize the memory path before the arithmetic.
• Prefer one self-contained artifact over a service graph when the problem allows it.
• Treat threat models and negative results as deliverables, not footnotes.

GitHub · SPEL · SPORE

Open to conversations on inference kernels, P2P transport, and ML infrastructure.