Skills

What I work with, demonstrated through code — all from AI OS.

  • Python

    Primary language. FastAPI backend, async background loops with error backoff, Pydantic models, threading locks, SQLite integration.

    → agent/services/agent_service.py — full message pipeline from user input to LLM response with context assembly, tool execution, and conversation persistence.

  • TypeScript

    Strict-typed React frontend. Interfaces, discriminated unions, custom hooks with reducers, async state management.

    → frontend/src/modules/workspace/hooks/useWorkspace.ts — full workspace state machine with file open/close, search, summarize, upload, all typed end to end.

  • SQL / SQLite

    Raw SQL schema design, FTS5 full-text search, WAL mode for concurrent reads, migrations, connection pooling, DB-backed virtual filesystem.

    → workspace/schema.py — virtual filesystem in SQLite with FTS5 indexing, chunking tables, and content-hash deduplication.

  • React

    Functional components, hooks, useReducer patterns, context, WebSocket integration, real-time UI updates.

    → frontend/src/modules/chat/ — real-time chat interface with WebSocket messaging, conversation sidebar, model selector, import wizard.

  • FastAPI

    REST APIs, WebSocket endpoints, middleware chains (CORS, GZip, HTTP logging), router composition, dependency injection, startup/shutdown lifecycle.

    → scripts/server.py — main application wiring all module routers, database mode switching, static file serving, subconscious wake on startup.

  • LLM Orchestration

    Multi-model support via Ollama. Prompt construction, context window management, streaming responses, model switching at runtime.

    → agent/agent.py — stateless LLM interface that receives assembled context and generates responses. No memory in the model; all state is external.

  • Embeddings & Semantic Search

    nomic-embed-text embeddings, cosine similarity scoring, keyword fallback when embeddings unavailable.

    → agent/threads/linking_core/scoring.py — embedding-based concept scoring with cosine similarity and spread activation.

  • Token Budgeting & Context Management

    Greedy packer that sorts facts by weight, fills top facts at requested detail level, downgrades tail facts to one-liners, drops overflow.

    → agent/threads/base.py — _budget_fill() method. Per-level budgets (L1=150, L2=400, L3=800 tokens). No thread can blow up the context window.

  • Tool Calling

    Text-native parser that finds :::execute::: blocks in LLM output, validates against a safety allowlist, executes, feeds results back for up to 5 rounds.

    → agent/threads/form/tools/scanner.py — the parser. Plus safety.py — the gate.

  • Cryptography

    Fernet symmetric encryption with machine-derived keys via PBKDF2HMAC. Encrypted at rest, auto-fallback if crypto library unavailable.

    → agent/core/secrets.py — full encrypted secrets store for API keys and OAuth tokens.

  • OAuth 2.0

    Multi-provider OAuth flows with token storage, refresh, callback handling. Gmail, GitHub device flow, Discord bot tokens.

    → Feeds/sources/ — each feed source has its own OAuth adapter with start, callback, status, and disconnect endpoints.

  • Event-Driven Architecture

    Event registry, priority levels (LOW → URGENT), handler dispatch, trigger system (time-based, event-based, threshold-based) with cooldowns and arm/disarm.

    → Feeds/events.py + agent/subconscious/triggers.py — feed events fire triggers that execute automations through tool adapters.

  • Systems Architecture

    Designed a multi-threaded context assembly engine with relevance scoring, associative linking (co-occurrence weighting, graph traversal, temporal decay), and hierarchical attention levels. Modular adapter pattern — every component is swappable.

    → agent/threads/linking_core/ — the relevance engine. Concept extraction, link consolidation, thread scoring for context assembly.

  • Fine-Tuning Pipelines

    Export per-thread training data to JSONL. Consolidation pass, per-thread export, combined dataset. MLX config management for Apple Silicon training.

    → finetune/api.py — neuroplasticity export pipeline. Conversations become training data that teaches smaller models to respect the OS's state format.

  • Testing

    392 pytest tests across 17 test files. Unit, integration, and behavioral. Fixtures, mocking, parametrized cases.

    → tests/ — run pytest tests/ -v, watch them all pass.

  • Docker

    Multi-stage builds (frontend builder → Python runtime), compose with volume management, Ollama host networking.

    → Dockerfile + docker-compose.yml — one command, clean build, production-ready.

  • CI/CD

    GitHub Actions running full test suite on every push and PR.

    → .github/workflows/ci.yml

  • Data Import & Parsing

    Multi-format parsers for ChatGPT, Claude, Gemini, VS Code Copilot conversation exports. vCard contact import from Google, iCloud, Outlook.

    → chat/parsers/ + import_contacts.py