YashDoke.

SkeletonGraph ↗ zero-LLM structural retrieval for coding agents · PyPI · MCP Registry

Coding agents waste most of their budget before they solve anything — grepping around a repository, reading whole files to find one function. SkeletonGraph replaces that search with structure: a tree-sitter symbol and call graph across ten languages, ranked by PageRank centrality, fused with BM25 and code embeddings through reciprocal rank fusion. No LLM calls in the index, so it rebuilds instantly after every edit.

Evaluated on SWE-bench Verified with Docker-verified test runs against eight competing pipelines, including aider's repo-map and two knowledge-graph tools. It matched a frontier agent's solve rate while cutting the expensive tail — which turned out to be the real finding: retrieval barely moves average cost, but it stops the worst 5% of tasks from running away.

HierMem ↗ constraint-preserving hierarchical memory for long conversations · PyPI

Long conversations decay in a specific way: the model doesn't forget everything evenly, it forgets the constraints you set at the start. HierMem borrows virtual-memory paging to fix that — a four-level hierarchy from a compact topic index down to raw turns, a stateless curator that decides what to page in, and a non-evictable zone for constraints that must survive.

Benchmarked over fifty-turn conversations across four domains against RAG, RAG+summary, MemGPT-style memory, and full-history prompting. The result that mattered wasn't the score — it was that placement policy, not retrieval volume, turned out to dominate whether a constraint survived.

PERCH a personal agent whose memory belongs to you, not to a vendor

Every assistant I use keeps its own private copy of who I am, and none of them talk to each other. PERCH inverts that: your context lives in one place you own — imported from the official exports of the assistants you already pay for — and is summoned by a hotkey inside whatever application you're already working in, answering beside your work and writing back into it.

The hard part was never the panel. It was the admission gate: a declarative, auditable rule layer over six user-owned memory classes that decides what any given model is allowed to see, and routes private classes to a local model instead of the cloud. A single global relevance threshold stops leaks but throws your identity away, so you stop sounding like yourself. Per-class floors keep both.

LLM Reasoning Pipeline ↗ diagnosing where reasoning breaks, then fixing that specific break

Most reasoning benchmarks tell you whether a model got the answer wrong. This one tells you which step went wrong. Ground-truth backtracking and error-propagation analysis locate the failing link in a chain, teacher–student diagnostics classify it, and then a QLoRA fine-tune targets that failure mode specifically rather than retraining on everything.

R2 autonomous stack ↗ ROS2 · perception · path planning under partial observability

I led software and AI for our Robocon robot: the ROS2 architecture, real-time perception, and path planning where the map is incomplete and the obstacles move. Validated in Isaac Sim and Gazebo before it ever touched hardware, then integrated with the embedded controllers — including closed-loop position-hold on an ESP32 mecanum base, so the robot actively resists being pushed off its mark.