rMax.ai
Enterprise AI Engineering

Projects

Infrastructure and tools for the agentic era.

Declarative conformance testing DSL for Model Context Protocol servers

A Hurl-inspired declarative DSL with a custom Python async runner for transport-agnostic MCP conformance testing. Write readable .mcph test files, run them against stdio or Streamable HTTP servers, and get CI-native JUnit/JSON reports.

Hurl-Inspired Syntax Readable, self-documenting .mcph test files for MCP server conformance with minimal ceremony.
Transport-Agnostic Native support for stdio subprocesses and Streamable HTTP transports in a single test file.
Rich Assertions JSONPath, regex, fuzzy type assertions, and variable capture with injection into subsequent requests.
CI-Native Output JUnit XML and JSON report formats for direct integration into CI/CD pipelines.

Production-oriented agentic systems PoC — governed AI workflow for house hunting

A full-stack house-hunting assistant that demonstrates governed AI workflow engineering: code handles determinism and constraints, agents handle qualitative judgment, and humans retain final authority through an approval gateway. The LLM is a narrowly-scoped component inside a conventional software system.

Three-Layer Architecture Code for determinism and constraints, agents for qualitative judgment, humans for authority over side effects.
Governed Workflow Engine State machine with explicit transitions, idempotency, retries, and full audit logging.
Approval Gateway All external side effects (emails, messages) require explicit human approval before execution.
Agent as Component LLM is a narrowly-scoped component; workflow, persistence, and security are conventional engineering.

Executable catalogue of 23 agentic design patterns using Google ADK

A practical reference implementation cataloguing 23 agentic design patterns — planning, reflection, tool-use, multi-agent, and governance — each as an isolated, runnable example built on Google's Agent Development Kit.

23 Runnable Patterns Isolated, executable examples of planning, reflection, tool-use, multi-agent, and governance patterns.
ADK-Native Each pattern leverages ADK's agent composition, tool binding, and session management primitives.
Progressive Complexity Patterns progress from basic tool-using agents to governed multi-agent workflows with structured state.
Pattern Catalogue CLI List, inspect, and run patterns interactively through a unified command-line interface.

Governed research runtime that turns questions into traceable claims

An enterprise-grade ADK research system that decomposes broad investigations into scoped questions, evidence collection, claim building, contradiction analysis, verification, approvals, and final reports.

Staged Epistemic Pipeline Moves from intent to scope, evidence, claims, contradictions, coverage analysis, drafts, verification, and approval.
Three-Plane Runtime Separates governance, deterministic workflow orchestration, and bounded cognitive agents.
Bounded Agent Roster Uses narrow-role agents for planning, evidence curation, contradiction search, verification, and synthesis.
Claim-First Quality Optimizes for supported, qualified, and traceable claims rather than polished but weak prose.

Reference implementation for loop engineering with Google ADK

A practical repository for building agentic loops with durable state, selective memory, bounded execution, verification, explicit progress tracking, and deterministic stopping conditions.

Three-Plane Architecture Separates deterministic control, ADK execution, and state and memory data flows so each concern stays legible.
Progressive Examples Demonstrates bounded document refinement, evidence-driven research, and resumable coding loops.
Deterministic Shell Models propose work, but budgets, approvals, and stopping decisions remain enforced in code.
Verification First Schema checks, lint, types, and tests run before any model judge can treat work as complete.

Adaptive agent harness foundry with deterministic configuration evolution

A HarnessX-inspired proof of concept that treats ADK harnesses as typed, versioned configurations, evaluates them deterministically, and evolves them through structured config patches rather than source-code rewriting.

Versioned Harness Configs Harnesses are first-class, hashed configuration objects with inspectable provenance and reversible evolution history.
Lifecycle Processors Typed processors attach to ADK lifecycle callbacks so behavior changes stay explicit and composable.
Trace to Patch Loop Structured SQLite and JSONL traces feed a meta-agent pipeline that proposes bounded configuration patches.
Deterministic Promotion Gate Candidate harnesses are promoted only after code-enforced validation and held-out benchmark evaluation.

Reference harness for bounded, verifiable autonomous software engineering

A coding-agent runtime that models work as constrained transitions over repository states, with a deterministic harness deciding whether to accept, reject, roll back, branch, or complete each proposed change.

State Machine Controller Enforces bounded execution and turns each proposed edit into an explicit transition decision.
Goal Contracts Typed YAML task bounds keep objectives, permissions, and stopping criteria explicit and machine-checkable.
Hidden Verification Evaluator plugins and protected tests separate proposal quality from self-reported model confidence.
Checkpointed Evidence Git-backed repository checkpoints and append-only artifacts make rollback, replay, and audit straightforward.

Scenario-driven test runner for evaluating MCP client and server implementations

A declarative conformance harness that runs HTTP and JSON-RPC scenarios against MCP servers, gateways, and client runtimes, then emits audit-grade pass/fail reports for protocol and auth behavior.

Declarative Scenarios YAML scenario sequences model OAuth, bearer-token, protocol, and error-path interactions step by step.
Partner Adapters The same conformance suite can target servers, gateways, or client runtimes through configurable adapters.
Protocol Assertions Checks HTTP status, headers, JSON bodies, JSON-RPC schema, and OAuth-specific response behavior.
Audit-Grade Reports Produces JSON and human-readable pass/fail output for certification, regression tracking, and side-by-side evaluation.

Spec-focused auth surfaces and a generic client CLI for end-to-end MCP auth testing

A FastAPI test server plus standalone mcp-auth CLI for discovering, authenticating, and exercising MCP resources across OAuth and bearer-token modes without ambiguity.

Explicit OAuth Resource /mcp/oauth accepts OAuth bearer tokens from auth-code, device, and client-credentials grants.
Deliberate Bearer Surface /mcp/bearer-token provides a static bearer-token contract isolated from OAuth discovery and issuance.
Generic mcp-auth CLI Resource-centric commands discover metadata, select auth mode, save profiles, and keep tokens valid.
Flow + Reference Docs Static docs provide exact walkthroughs and endpoint behavior constraints for reproducible client validation.

A governed knowledge layer for enterprise agents — ontology, policy, approval, provenance

A complete, runnable Python research repository demonstrating how enterprise agents can safely move from information retrieval to governed SaaS write operations — through operational ontology, deterministic policy evaluation, payload-bound approval, and hash-chained provenance.

Operational Ontology RDF/Turtle + SHACL ontology of actors, roles, actions, policies, evidence, approvals, and audit records.
Deterministic Policy Engine YAML-defined policies evaluated in code with zero LLM dependency for enforcement decisions.
Payload-Bound Approval SHA-256 integrity verification ensures approved payloads match executed operations exactly.
Provenance Ledger Hash-chained audit events with 25 competency scenarios and comparative evaluation across capability levels.

Decoupled vector search prototype with immutable index artifacts

Shardlake is a Rust prototype for personal-scale vector search inspired by billion-scale architectures. It separates raw embeddings from serving, builds sharded indexes offline, publishes immutable manifest-driven artifacts, and answers queries through stateless HTTP workers with lazy shard loading.

Decoupled Data Plane Raw vectors are ingested into versioned storage separately from query-serving indexes and aliases.
Offline Shard Builds K-means sharding and shard-local search produce reproducible index artifacts without coupling build time to serving time.
Lazy Query Serving Stateless HTTP workers route via persisted metadata and only load shard bodies on demand to keep memory bounded.
Benchmarkable Lifecycle Manifest publishing, aliases, and a recall and latency benchmark harness make versions easy to compare and audit.

Technical book + research artifact on harness-first, governed AI software engineering.

A working book and repo exploring AI-first software engineering where the harness is the primary design surface. Includes patterns for governance, evaluation/traces, and memory systems, plus a deterministic kernel loop (planner/writer/critic) and eval contracts to keep autonomy safe and reproducible.

Harness-First Design Treats the harness as the primary design surface for reliability, cost, and iteration speed.
Governance + Tracing Patterns for policy, evaluation, and traces to make autonomy auditable and safe.
Deterministic Kernel Loop A planner/writer/critic loop designed for reproducible behavior under constraints.
Eval Contracts Contracts and checks that keep autonomy bounded, testable, and replayable.
RX

Minimal autonomous systems agent with a microkernel architecture

RX is a goal-directed execution engine capable of modifying files, running commands, invoking tools, and operating locally or in distributed mode. It is not a chatbot: the kernel owns reasoning and iteration, tools own side effects, state is append-only, and transport is replaceable.

Kernel Decides Executes the autonomous loop, dispatches tool calls, enforces iteration limits, and evaluates termination.
Tools Act Filesystem, shell, and networking side effects live in the tool runtime, not the kernel.
State Persists Structured events are persisted append-only (memory today, sqlite next), enabling replay and auditability.
Transport Delivers CLI, HTTP, and workers are interchangeable front-ends; transport can evolve without kernel bloat.

Budget-Literate Autonomy for Agentic Systems

Ratelord provides a "sensory organ" and "prefrontal cortex" for resource availability and budget planning, enabling autonomous systems to negotiate, forecast, govern, and adapt under constraints.

Local-First Architecture Zero-ops daemon that runs alongside your agents, providing low-latency constraint management.
Predictive Modeling Time-to-exhaustion forecasts help agents plan activities based on real-time resource availability.
Intent Negotiation Standardized protocol for agents to request resources and negotiate budgets before execution.
Event-Sourced Audit Complete audit trail of all decisions and allocations, fully replayable for debugging and analysis.

The authoritative control plane and global edge runtime for the rmax.to redirect fabric

A typed, schema-first redirect system managed via a CLI-first workflow, treated as critical identity infrastructure rather than a generic marketing tool.

Global Edge Runtime Stateless Cloudflare Workers serving from KV at the edge with <10ms global latency.
Safety Guarantees Strict enforcement of invariants: no loops, reserved names, and disable-over-delete policy.
CLI-First Workflow The rlinks CLI is the only way to mutate state, enforcing a strict validation pipeline.
Typed Redirect Objects Redirects move through a strict lifecycle with forward-only versioning and audit trails.

Smart pointers for hermetic binary management

A repository of DotSlash wrapper files and tools that facilitate fetching, verifying, and running executables transparently across platforms.

Polyglot Manifests Tiny JSON files with shebangs that point to large binaries, avoiding "bloat" in git repositories.
Content Addressing All binaries are verified by hash (sha256/blake3) before execution, ensuring integrity.
Cross-Platform shim Automatically detects OS/Arch to fetch and run the correct binary for the current host.
Hermetic Sandbox Include scripts for running manifests in secure, isolated Podman containers.