- 5
- Projects
- 1
- In production
- 1
- Delivered
- 3
- Research
- 2,306
- Tests green
Production
LAS — Life Assistant
Reeb Labs's own production AI — a second brain that learns, remembers, and acts. The discipline we sell, proven on ourselves daily.
- Problem
General LLM products forget every conversation, can’t act on private data, and treat memory as transient prompt engineering. Anyone serious about a “second brain” needs persistent context, autonomous workflows, and reasoning over their own history.
- Approach
Temporal knowledge graph (Graphiti) over a hybrid retrieval stack — dense embeddings, personalized PageRank, Cypher consolidation. Five-agent cognitive hierarchy with intent routing, decision journaling, and meta-cognition. Local inference on oMLX (Qwen 3.6 35B-A3B for entity extraction, Apple Silicon), Anthropic primary for hard reasoning with deterministic local fallback. Phase 0.5 KG re-architecture validated against a 2540-line spec with a ≤15h rollback budget.
- Outcome
In daily production for one user. 2306 hermetic tests green. Phase II-B deployed (hybrid embeddings, PPR, Cypher consolidation). Phase III queued — Thompson Sampling, self-healing, trust metric — gated on 14-day stability and r ≥ 0.6 correlation with human judgment.
GreenScan
Industrial sustainability assessment, built end-to-end and shipped as a self-contained product to a single founder.
- Problem
A founder needed a turnkey scanning and reporting tool for sustainability metrics, with no appetite for ongoing platform dependence. They wanted to own the codebase outright.
- Approach
Full stack — data ingestion, scoring engine, PDF reporting — documented for handoff. No SaaS lock-in. No recurring billing. Knowledge transferred alongside the code.
- Outcome
Delivered. Operating independently of us by design. We treat clean handoffs as a deliverable, not a failure mode.
Active research programme
ARGUS
Market intelligence engine for SMB operators — scanning, briefing, and reporting without analyst headcount or per-report SaaS pricing.
- Problem
Owner-operators need fast, defensible competitive scans, but most “AI competitive intelligence” tools are thin wrappers around web search or charge per-output. Internal analysts are out of reach for SMB budgets.
- Approach
Software-only — scrapers, structured extraction, embeddings, scheduled briefings. Local-first inference via oMLX. Scheduling and orchestration built on the same primitives as LAS. Distinct from Sentinel, which is a hardware program.
- Outcome
Architecture validated on prototype data. In development against a sharper SMB design partner.
ATHENA
A single product being unified out of two earlier R&D tracks (athena and athena2). The pause is intentional — convergence is a design question, not a refactor.
- Problem
Two parallel prototypes were exploring different angles of the same applied AI problem. Running them side-by-side was clarifying. Shipping them as one is what makes them a product.
- Approach
Treating the merge as a first-principles design pass: which interfaces survive, which were artifacts of the parallel exploration, what the unified data model looks like. No code merged blindly.
- Outcome
Architecture being unified. Resumes when ARGUS unpause window opens — both share substantial infrastructure.
Sentinel
A drone hardware program, distinct from any software-only scanning work. Some applied-AI problems require physical sensors, not just data ingestion.
- Problem
Industrial inspection and perimeter awareness need actual sensors in the field. Off-the-shelf drones don’t expose enough of their pipeline to plug cleanly into a custom AI stack.
- Approach
Custom hardware build — parts sourced and prototyped in Brescia. Software stack designed against the same inference and orchestration layer the rest of the lab uses.
- Outcome
Stale since February 2026. Components held. Restart conditional on a credible industrial pilot — we don’t run hardware programs on speculation.
What this list excludes
We don’t list internal tooling (oMLX, FLOS, sync infrastructure), exploratory R&D that hasn’t earned a product name yet, or work covered by client confidentiality. The five above are the projects we’re willing to be measured against.