software

Tight upper and lower bounds on how far chain-of-thought carries a transformer before tool delegation becomes necessary — with explicit constants.

Workflow-atomic GPU-cluster scheduler for AI agents — within 1.31× of Bélády-optimal KV-cache eviction, with OpenMP-accelerated kernels and LangChain / AutoGen / CrewAI bridges.

Computable bounds on knowledge-editing side effects, plus the impossibility result ruling out perfect locality and generalization at once.

A three-stage verify-then-ground pipeline for financial document QA that detects and grounds hallucinated claims.

Conformal-coverage router for cost-aware LLM cascade serving.

DAG-structured, step-level evaluation harness for multi-step agents, with error-propagation tracking.

Knowledge-graph-augmented retrieval for multi-framework regulatory gap detection.

The program site — five post-Transformer candidates evaluated head to head, with the orthogonality argument for why each targets a different ceiling. Browse the live showcase.

Per-token learned routing across SSM, sliding-window, and full-attention mixers.

Tokenizer-free, byte-level language model — Hierarchical Entropy-Linked Information eXchange.

Block-diffusion language model — parallel multi-token decoding within each block.

Test-time neural memory — a memory module that learns and updates during inference, not just training.

Continuous-thought reasoning — thinks in latent space and allocates its own thinking budget per token.

An agentic harness with citation-grade memory: every fact carries a source URI, a temporal validity window, and an append-only provenance ledger. MCP-native, multi-frontend, fully self-hostable.

Verifiable memory for LLM agents — every recalled claim is HMAC-signed back to its originating trajectory span.

Turns any OpenAPI 3.x spec into a Model Context Protocol server, with auth, rate-limiting, and OpenTelemetry built in.

A Postgres MCP server for agents with layered safety — role grants, a pglast AST guard, per-transaction envelopes, audit logging, and pgvector. PG 13–17.

A research MCP server composing arXiv, Semantic Scholar, and OpenAlex — related work, citation graphs, and BibTeX in the chat.

An MCP server that hands coding agents live Jupyter kernel state — variables, dataframe summaries, plots, tracebacks — not just the notebook JSON.

OpenTelemetry-native, local-first observability for AI agents — DuckDB on disk, a localhost viewer, and adapters for Anthropic, OpenAI, LangGraph, AutoGen, and CrewAI.

Fifteen reusable, agent-agnostic browser recipes plus an MCP server, so browser-using agents stop re-discovering cookie banners and infinite scroll.

Probabilistic certified robustness and anomaly detection against knowledge-graph poisoning in retrieval-augmented generation.

Distribution-free, spatially adaptive uncertainty quantification for neural-operator PDE surrogates via physics-informed conformal prediction.

A temporal-logic framework for formal verification and repair of LLM-generated business-process models.

Safety-preserving continual domain adaptation of LLMs via Fisher-based subspace identification and orthogonal gradient projection.

A library of probes for agent benchmarks — contamination, gold-answer leaks, harness-injection, and reward hacking, with confidence intervals on every result.

Audits AI-agent benchmarks for the eight exploit families that quietly inflate reported scores.

A small, reproducible benchmark for RAG pipelines.

An open-source benchmark for Claude Code skill bundles and CLAUDE.md configs — pass@k, cost, and reliability.

A rigorously verified map of the theoretical foundations of LLM reasoning — transformer expressivity, chain-of-thought error bounds, circuit complexity, logical characterizations, learnability.

An assumption-explicit catalog of published impossibility and lower-bound results for LLMs and agents — circuit-complexity ceilings, hallucination bounds, watermarking, alignment.

A theory-first map of why reasoning models (o1/o3, DeepSeek-R1, Claude-thinking, QwQ) actually work — chapters, annotated papers, model comparisons, and reproduction notebooks.

An interactive atlas of discovered circuits and sparse-autoencoder features in LLMs, with Colab reproductions on open-weights models.

Watch an LLM think — sparse-autoencoder features firing live across every token, on a laptop, no GPU required.

A production prompt workflow — semantic diff, eval-on-PR, lockfile, drift detection, and rollback for markdown prompts in a repo.

A reproducible catalog of LLM behaviors that vanished as models scaled.

Local semantic code search — a CLI and MCP server that run entirely on your machine.