Templet Solutions builds production software — distributed systems, infrastructure, and full applications — alongside original research in geometric deep learning. Different problems, same standard: built right, not just shipped.
Four papers · two shipped systems
A sequence of papers developing a geometric and field-theoretic framework for AI memory architecture — from discrete graph theory through scalar-tensor gravity on hyperbolic belief manifolds.
Introduces a discrete geometric framework for knowledge graphs in which beliefs carry a local metric tensor, contradictions manifest as negative Ollivier–Ricci curvature, and information homeostasis emerges from the balance of a curvature-smoothing flow and epistatic regulation.
Instantiates the RBM framework in the Poincaré disk with zone-specific inertia values encoding epistemic authority. Develops the HSH metric, the curvature-adaptive signed Laplacian, and the coherence field as a memory commit criterion.
Presents VeritasMemoria's three-layer defense architecture: immutable Task Registry, three-gate memory validation with hyperbolic quarantine, and a five-domain goal stability invariant. Empirical evaluation across 3000+ simulated attacks demonstrates 98–100% detection for overt attacks and zero successful hijacking.
Derives a closed-form scalar-tensor field theory from the HSH architecture. Proves via the Bianchi identity that contradiction stress is conserved and can only be removed by human intervention — a field-theoretic derivation of the system's architectural human-oversight requirement.
Production systems built to solve specific problems in high-stakes domains.
Local-first AI memory architecture for legal practices, healthcare providers, and independent professionals. Four-zone belief graph with cryptographic audit trails, human-gated contradiction resolution, and the geometric validation layer developed in the research program.
A project manager and story bible for fiction writers. Reads a manuscript and builds a belief graph of its world — characters, continuity, plot threads, and knowledge asymmetry — so writers can query their own story instead of re-reading it.
Lead intelligence pipeline with waterfall-style source prioritization — cheapest/fastest sources first, escalating to richer sources only when needed. Includes idempotency layer, rate limiter, execution engine with producer/worker separation, and doubt-scored signal quality assessment.
Code that exists to be useful, not to be impressive.
Distributed rate limiting service implementing a Sliding Window Log algorithm via Lua scripting for strict atomicity.
High-reliability asynchronous task queue built with Python and Redis.
Redis-backed idempotency implementation that prevents duplicate processing in distributed systems.
High-integrity activity tracking system for compliance and observability, capturing a complete history of system changes using PostgreSQL.
Doubt-score-gated lead enrichment pipeline. Built in both Rust and Python.
Owner finance property finder.
Reference implementations of fundamental systems programming building blocks, written from scratch against the standard library with no external dependencies. Built in both Rust and Python.
Minimal Redis-compatible in-memory key-value store speaking RESP over TCP. Built in both Rust and Python.
A JSON parser written from scratch, without serde_json. Built in both Rust and Python.
A multi-threaded, memory-mapped grep clone. Built in both Rust and Python.
A hyper-fast binary semantic grep command-line search utility. Built in both Rust and Python.
Lightweight, network-accessible, in-memory key-value store speaking a Redis-compatible subset of RESP over TCP. Standard library only.
Concurrent CLI link checker and asset crawler — worker pool, context-based cancellation, token-bucket rate limiting, robots.txt support. Standard library only.
Agent that writes journal entries reliably.
MCP server giving Claude direct access to GitHub — repos, files, issues, PRs, and Actions — via a personal access token.