Research BuildsDense Retrieval ResearchManuscript draft in progress
Local-first retrieval research system for testing whether post-hoc vector-space corrections can improve dense retrieval without retraining the base embedding model. The build emphasizes reproducible corpus evaluation, failure analysis, metric comparison, and saved experiment artifacts so retrieval improvements can be measured across iterations.
Technical angle: Dense and sparse baselines, BM25/dense blending, reciprocal rank fusion, query rewrite branches, HyDE expansion, CPU reranking, checkpointed execution, Parquet/JSON artifacts, notebook validation inputs, and Recall@K/MRR/latency comparison.
Research BuildsModel Architecture Research
Research-grade semantic mixture-of-experts architecture for small-compute environments. The implementation combines a tiny transformer backbone, token-level expert MLP dispatch, a MeaningSpace teacher router, a learnable student router, hybrid routing with load penalties, and centroid-based clustering so specialization can be studied without assuming large-scale infrastructure.
Technical angle: Tiny transformer backbone, MoE dispatch layer, semantic teacher routing over cluster centroids, student-router distillation, hybrid score blending, load-balancing penalties, centroid updates, reclustering utilities, routing traces, metric aggregation, snapshot persistence, and an end-to-end training loop across core, routing, clustering, and observability modules.
Research BuildsAI Operations Planning System
Local-first planning and AI-operations system that turns scoped goals, retained feedback, and iterative model output into progressively sharper execution plans. The current build adds routing, prompt governance, guardrails, and local evaluation around the planning loop.
Personal motivation: if AI can assist with routine execution, it should also help with the heavier cognitive load around ambiguous problems: repeated search, planning, refinement, and decision pressure. This project explores an AI-assisted planning loop that externalizes that burden into structured iterations, retained feedback, and reviewable next steps.
Technical angle: Feedback-driven planning iteration, prompt registry, model-call capture, guardrail validation, local evaluation evidence, scheduler controls, run-history persistence, and operator-readable execution artifacts.
Research BuildsSimulation and Analytics
Local-first artificial-life simulation platform that treats evolution as a data and systems problem instead of a visual toy. The stack separates a FastAPI inspection surface, a simulation worker, and a shared worldsim domain layer that persists runs, ticks, host states, lineage summaries, world patches, birth/death events, interaction events, and interventions.
Technical angle: SQL-backed simulation runs, deterministic tick advancement, seeded world bootstrapping, host and lineage state tables, birth/death/interaction event persistence, mutation and phenotype decoding, intervention APIs, replay endpoints, live WebSocket payloads, world-map inspection, lineage detail routes, and dashboard-ready telemetry for population, resources, births, deaths, lineage count, and alive score.
Research BuildsCognitive Control Loop Research
Local modular daemon prototype for studying continuous cognitive-style control loops with sensor-driven thought cycles, adaptive memory routing, structured event persistence, deterministic simulation, and API/CLI inspection surfaces.
Technical angle: Explicit event modeling, resource-aware memory routing, deterministic signal simulation, persisted loop output, FastAPI health/event endpoints, CLI-driven local experiments, and repeatable inspection of self-prompting behavior.