HEROS-RL
Critic-free GRPO for zero-downtime PostgreSQL migrations: deterministic format, SQL safety, and forge rewards; 500-case dataset.
33 tests · toy G=4–32 solves task · 1.5B run pending
itsoumya-d · applied AI systems
Agent infrastructure,
measured in public.
Ten systems spanning RL policy optimization, agent security, retrieval, kernel optimization, post-quantum transport, routing, and governance. Evidence links to source, green CI, and local benchmarks; claims that need unavailable GPUs or credentials are marked pending.
Reflexion-SWE
Actor/evaluator/self-reflection loop with AST-tested code repair and training-free GRPO verbal priors.
40 tests · toy: base 0.00 → full 1.00 · model run pending
GraphRAG-CodeAudit
Local-to-global hierarchical GraphRAG over code repositories, with community reports and a flat-RAG baseline.
15 tests · 100 questions: GraphRAG 7.58 vs flat 7.07 · 64% wins
HEROS-Firewall
Inline MCP security proxy for OWASP ASI01–ASI10, HMAC human approvals, and Merkle audit receipts.
16 tests · 1,000 adversarial / 200 benign · FNR 0, FPR 0 · p99 0.103 ms
RouteLLM-Gateway
Calibrated Bradley–Terry routing with semantic cache and OpenAI/Anthropic-compatible streaming.
12 tests · 98.38% retention · 69.40% cost reduction (synthetic)
Anthropic-PerfBench
NEON and WGSL transformer kernels across three optimization iterations with a memory roofline.
RMSNorm 3.09× · RoPE 2.35× · INT4 GEMV 11.83× vs scalar
edgeinfer v2
Browser-local ONNX inference, INT8 Matryoshka retrieval, and gamma=5 speculative decoding.
95 tests · 100k chunks p50 25.1 ms · Recall@10 0.725→1.000
peervault
Post-quantum WebRTC transport: FIPS 203 ML-KEM-768 + P-256 hybrid handshake and encrypted data mesh.
100/100 NIST KAT · 75 tests · encrypt 1395 MB/s local
Mobbin-UI-Compiler
Reference-driven compiler for five SaaS surfaces with accessibility-invariant source tests.
10 tests · 5 surfaces · live Mobbin browser audit pending
HEROS-Cloud-SaaS
PLLB governance studio: human-approved natural-language policies and per-step Merkle receipts.
6 tests · 70% → 100% completion · 4 policies · synthetic simulation