Sep 10, 2026
Extension for Pi coding agents that fuses actions, packs observations, and compacts context to cut token waste
SoL-Pi layers four opt-in mechanisms onto unmodified Pi releases via public extension APIs: action fusion for combined edit-and-validate calls, stable observation handles with paged recall, evidence-preserving log reducers, and pressure-aware compaction of completed steps. These targeted reductions of repeated model turns and oversized context are useful only to existing Pi users running long agent sessions and do not generalize beyond that harness.
Server for storing querying reviewing and sharing hierarchical agent trajectories with multi-format imports
It runs a SQLite-backed store behind a REST API and MCP server that ingests traces from OTLP JSON Langfuse Tardigrade and similar sources then serves activity trees analysis tables and shareable views. The approach follows the pattern of established LLM observability tools yet targets the narrower audience of teams that need legal-benchmark workflows and deterministic hierarchical review rather than general-purpose tracing.
Operator-based pipelines for LLM data synthesis, cleaning, and filtering with agentic assembly
DataFlow supplies reusable operators built from rules, models, and LLM calls that are composed into pipelines for synthesis, extraction, evaluation, and filtering across domains such as healthcare and finance, plus agents that dynamically recombine or invent operators on demand. The design follows conventional pipeline and ETL patterns yet adds LLM-centric operators and agent-driven construction, targeting teams that need reproducible high-quality training data for domain-specific models rather than general-purpose ETL users.
Estimates dense 3D contact maps and force distributions on hand and articulated object meshes from monocular egocentric RGB plus object geometry
The three-stage pipeline extracts cross-modal visual and geometric features, refines object pose through hand-mesh registration, and feeds the fused representation into an InteractionGNN that outputs per-vertex contact and 3D force on MANO and object meshes. The approach is the first to move force estimation beyond image-space or planar assumptions by training on physics-simulated per-vertex supervision, so adoption will remain confined to researchers working on egocentric AR or robotic manipulation rather than broader developer audiences.