stanley_tejakusuma

A skill that turns any transcript into a structured, confidence-tagged brief. Point it at a local file or a link from YouTube, TikTok, X and podcasts alike, and it fetches captions or transcribes locally before digesting. It distrusts diarization on purpose: attributes lines by content instead of a speaker label that's often silently wrong, and tags every claim CONFIRMED, LIKELY and UNVERIFIED instead of pretending uniform certainty.

A read-only skill that renders Things 3 board state into an agent's context: the full board for planning or a lighter search-first pass before creating a task. It only ever observes. Proposing or writing a change is a different skill's job entirely.

Reconciles a Things 3 board against a conversation: scans for task-shaped statements and produces an explicit add, update, move or close diff before touching anything. A WIP cap, a required blocker note on anything Waiting, and a closure needs real evidence, not a vibe.

A read-only CLI for a Telegram archive in TimescaleDB: search, freshness checks, message context, link extraction, CSV and JSON export, all shipped over SSH and docker exec so the database never needs direct exposure. Ships the query half only. The ingestion half needs a Telegram session that's effectively full account access, and that's not something to copy-paste from a stranger's example.

An MCP server that gives any agent access to your Oura Ring data: sleep, readiness, and activity as clean tools over the official v2 API. Bring your own token. Built so an assistant can reason about your recovery instead of just charting it.

A self-hosted agentic RAG gateway on the Hermes runtime, operated daily: multi-source ingestion across chat platforms, X watchlists, Reddit, and call transcripts, a tool registry, scheduled agents, and self-hosted web search. Any action that reaches the real world routes through human approval first.

A tmux-backed session manager for the Pi coding agent: /bg hands a running session to tmux and frees the terminal. The dashboard lists every session on the machine to reattach with a keypress. Reattaching resumes the same process, not a transcript replayed into a new one. Liveness is verified by process identity, so a crashed session disappears rather than lying about its state.

An architecture write-up for capital-bearing LLM agents: split the system into a brain that can only propose and a deterministic signer that verifies every intent through a layered fail-closed pipeline. No LLM in the signing path, ever. The pattern behind my live agent execution, published as docs and diagrams.

A fail-closed gate between an agent harness and irreversible actions: an LLM classifier with human confirmation, Ed25519-signed single-use intents bound to the exact action, and a red-team harness holding recall 1.0 across 59 adversarial cases. Ships with a shadow-runner path for deploying it without breaking anyone.

A self-hosted personal-finance stack: a dashboard for net worth and runway with an interactive spend-sensitivity slider, a REST API with audit logging, and ingestion via a Telegram bot and iOS Shortcuts. Built end to end in two days, then trimmed to what real use proved necessary.

An agent that sits on top of Uptime Kuma: batches alerts, diagnoses the failing container, and proposes (or applies, with approval) the fix over Telegram. Ops turned into a conversation.

Controlled fine-tuning experiments toward a self-hosted LLM: QLoRA on 14B and 32B bases, SFT then DPO, graded by a 711-case eval harness across 14 categories with a judge ensemble on high-stakes cases. The final version beat the hosted baseline on the harness, then lost a live A/B zero topics to thirty-five and hit the pre-registered kill rule, so the fine-tunes were shelved. The harness is the point. A measured negative result is still a result.

A post-trade journal that runs itself: a Telegram bot watches my futures account, pings me for the rationale while the trade is fresh, and a language model classifies the trading psychology across the full event timeline once the trade closes. Turns hindsight into a structured behavioral dataset.

A betting-strategy lab: 79 scripted strategies across dice, blackjack, roulette, baccarat, and more, from Martingale to D'Alembert, built to measure exactly how every progression system meets the house edge. Negative EV, empirically confirmed. The orchestration, bankroll simulation, and variance tracking outlived the thesis.

A projection system that merges four separate memory sources into one lint-enforced knowledge graph: around 6,000 atomic nodes and 21,000 semantic edges built from embedding similarity, with conflict-aware sweeps and annotation-preserving regeneration. It powers retrieval for my agents and doubled as the training corpus for atlas. Basically: one place where nothing I've learned gets lost or contradicts itself.

A full-stack decision engine: Hi-Lo counting, basic strategy with index deviations, and Kelly bet sizing, on a pure-logic core with zero framework dependencies. Applied probability you can drill with.

A TradingView indicator that models price dispersion with probabilistic standard-deviation bands. Gives a statistical read on volatility and likely ranges instead of fixed levels. I got tired of drawing horizontal lines and pretending they meant something.

A self-hosted market-data pipeline that backfills and incrementally archives one-minute OHLCV for 388 instruments, 223 crypto and 165 US equities, into a TimescaleDB hypertable, served behind one canonical FastAPI read API for both asset classes. Built for reliability, with freshness watchdogs, timezone-aware coverage checks, and multi-route egress.

A self-hosted pipeline that archives the Telegram channels I follow into TimescaleDB. Incremental sync, freshness watchdogs, and a search API over years of message history. Built because scroll-back is not a database.

A workspace for researching, backtesting, and deploying multiple quantitative strategies under one engineering process. Built on vectorbt and CCXT, with a per-strategy spec-and-review gate so every program is reproducible and comparable before it touches risk. If it can't survive review, it doesn't ship.

An autonomous execution-and-risk layer for a live futures signal feed. It parses incoming signals, applies adaptive risk controls, and executes with slippage-aware entry logic. The engineering is in the risk management and trade reconciliation, not the signals.

A NAS-based homelab that runs my life's backend: market-data and Telegram archivers, dashboards, bots, and monitoring as a Docker fleet, with backups, alerting, and multi-network egress. Uptime is a feature.