I turn frontier model
capabilities into products people can actually use.

Aligned listens across the surfaces work actually happens on (Zoom, Meet, Teams, Slack, Gmail, Calendar), then runs every transcript through an AI pipeline that extracts decisions, commitments, and ownership into a diff-style review UI where every human edit is captured.

The shape of the system:

• Company Brain. A pgvector memory layer that reads across meetings, so Aligned can say “you decided this three sprints ago” before the debate starts over.
• Agent frameworks in production. A CrewAI sprint planner for capacity-aware prioritization, rebuilt on a user-scoped Mem0 memory layer: CrewAI's default memory is process-local with no tenant partitioning, which leaks one customer's retrieval context into the next on shared FastAPI workers. CI greps every commit so nobody flips it back. Alongside it, a streaming email-triage pipeline and computer-use agents (OpenClaw + VNC) that drive full Ubuntu desktops: terminal, files, any GUI app.
• Real-time voice agent. A LiveKit speech-to-speech agent (xAI Grok primary, OpenAI fallback) exposes ~60 backend tools, with an optional LemonSlice video-avatar integration. A fail-closed, multi-layer HMAC auth chain with nonce replay protection and a heartbeat watchdog (covering 75 of 86 voice routes) means a leaked service token alone can't impersonate a user.
• Bidirectional ticket sync. Jira, Linear, Asana, Monday, Notion, and GitHub: per-provider mapping for project, status, priority, and issue-type, and a status that only reads “synced” when zero errors fired. Round-trip sync is 5× the work of write-only, but the customer keeps their tool of choice and Aligned becomes the connective tissue underneath.
• SOC 2 CC7-style controls (not a certification). SSO-only, per-tenant isolation, AES-256-GCM versioned token encryption with key rotation, AWS Secrets Manager for per-tenant webhook secrets, hash-chain audit logs, fail-closed rate limiting, and dual TS+Python environment validators that block startup on missing or malformed config.

Glassbox puts a fixed crew (planner, coordinator, four workers, validator, improver) on a live tldraw board fed by a Redis event stream, so every decompose, handoff, and verify lands on screen the moment it happens. The same crew runs two ways: as live Claude Code or Codex sessions you supervise from the command center, or headless against a graded benchmark.

The shape of the system:

• One engine, eight loop shapes. The engine never changes (decompose, dispatch, verify). A loop is that engine plus a stop condition: Land stops when the goal verifies, Climb when a metric plateaus, Sweep when a backlog drains, Race when a judge picks a winner.
• No fake progress. Workers author each edit with W&B Inference against the validator's failing cases and keep a change only if the built artifact scores higher. Agent Mail carries the handoffs, Beads tracks the dependency graph, and nothing advances on a timer.
• It improves itself. The benchmark ports a BPE tokenizer to Rust, scored by an exact token-ID diff against tiktoken. Between versions, the improver reads the failing cases back from Weave and rewrites the planner skill, so accuracy climbs without anyone touching the swarm code.
• Point it at your own repo. A task is just a goal, a workspace, and a checkable evaluator. Give Glassbox a repo and a test command and it clones into a disposable sandbox, finds the failing tests, and fixes them. Your source is never touched.