Hi, I'm Jason Graham and I did not write this. This is the front door to the lab: part notebook, part command center, part suspiciously glowing server rack.
I am interested in systems that make people more capable: agentic developer tooling, durable memory, secure execution, data platforms, retrieval, graph-shaped knowledge, and boring old databases that quietly do impossible things before breakfast.
My starred-repo constellation says the same thing in a louder accent:
agent runtimes sandbox boundaries local AI infrastructure
knowledge graphs RAG + memory embedded/vector databases
Rust + Go + Python SQL performance terminal-native workflows
security research model routing tools that feel like powers
|
|
I sampled my GitHub stars and the signal is not subtle:
| Orbit | Repos in the gravity well | Why it matters |
|---|---|---|
| Agentic engineering | goose, multica, everything-claude-code, agent-toolkit, sipeed/picoclaw |
Practical AI agents need harnesses, skills, orchestration, review loops, and boring reliability. |
| Memory + knowledge | cognee, claude-mem, SimpleMem, graphify, RAGatouille |
Context is infrastructure. Retrieval is not a feature; it is a nervous system. |
| Sandboxes + security | fence, agent-safehouse, nono, pydantic/monty, CyberGym |
If agents can run tools, tool execution needs walls, receipts, and blast-radius thinking. |
| Databases + search | kuzu, USearch, zvec, libmdbx, HaloDB, azimutt |
The future still depends on indexes, storage engines, query plans, and schemas that make sense. |
| Terminal sorcery | rtk, repomix, vim-dadbod, klaw.sh, llmfit |
The command line remains undefeated when it is fast, composable, and legible. |
| Data + finance trails | InsiderTrader, tablerag, sfquickstarts, Data-engineering-with-dbt |
Structured data is where the receipts go when they do not want to be found. |
|
Tracing cryptoscams for fun and profit. Follow the money, model the graph, keep the flashlight steady. |
Monthly Tabular Kaggle Challenges. Structured data, experiments, notebooks, and the eternal argument with feature engineering. |
|
Notes and experiments around language, models, representations, and the machinery behind meaning. |
Protohackers exercises in Elixir. Protocol puzzles, concurrency reps, and small distributed-system bruises. |
|
A public surface for thinking about data, systems, and how to make useful things from noisy inputs. |
Competition notebooks, experiments, and the archaeology of trying many things until the leaderboard blinks. |
mindmap
root((Jason))
Agentic Systems
skills
orchestration
memory
tool use
evals
Data
notebooks
tabular modeling
SQL
dbt
graph analysis
Infrastructure
local-first
sandboxes
observability
reproducibility
Interfaces
terminal
docs
dashboards
small useful apps
01. Build the smallest thing that reveals the truth.
02. Keep receipts: logs, tests, notes, commits, traces.
03. Prefer tools that compose.
04. Treat data as evidence, not decoration.
05. Make agents earn trust through constraints and verification.
06. Optimize for loops: learn, instrument, revise, repeat.
07. If the system cannot explain itself, add observability before mythology.
08. Fast is good. Correct is better. Fast and correct is a portal.
- AI agents moving from chat boxes into tool-using operating environments.
- Local-first personal AI infrastructure: private memory, private indexes, private automation.
- Capability-based sandboxes for running generated code without turning the laptop into a sacrifice.
- Embedded graph and vector databases as the new pocket engines.
- SQL performance traps, index-defeating expressions, and query plans with hidden knives.
- Data products that make uncertainty visible instead of hiding it under a glossy chart.
I am usually happiest around:
- practical AI tooling
- data systems
- weird datasets
- fraud trails
- developer automation
- security-minded agent workflows
- notebooks that turn into durable systems
- small tools with unreasonable leverage
If that overlaps with what you are building, wander into the repos and leave a signal.