The Lab

Experiments before they become products.

Curiosity in public — hypotheses, thin prototypes, and what each attempt taught. Not a roadmap of invented startups.

Currently exploring

Where attention is going.

Experiments

Shipping · private tool

Hypothesis: A private admin on the same Worker can keep a knowledge library honest and current without turning the site into a generic CMS product.
Prototype: GitHub-allowlisted /admin, D1 documents, R2 media/overlays, first-party analytics overview, publish rebuilds for search artifacts.
Technology: Workers · D1 · R2 · KV sessions · Analytics Engine + D1 event mirror
Learning: Scope stays “this site only.” Multi-tenant CMS fantasies can wait; publishing cadence cannot.

Planning · private workspaces

Hypothesis: Decade-scale products need full planning (users, architecture, risks, decisions) before production code — and that planning must stay private until Prototype.
Prototype: CMS Incubator module with 20-section workspaces; seeded eClearBag, ScholarQuest next chapter, UKSathi, NepalIPMS Public Demo.
Technology: D1 planning tables · admin UI · public index gated to Prototype+
Learning: Public visitors should never see unfinished internal research. Plan like a company; ship when ready.

AI Agents

Hypothesis: An agent that retrieves from a firm’s own matters and register can draft useful next steps — if every answer cites a source and never auto-files.
Prototype: Retrieval over matter notes + trademark FTS, draft outline with citations, forced human confirm before anything touches a client file.
Technology: Workers · embeddings API · D1 FTS · review queue patterns from NepalIPMS
Learning: Agents fail when scope is open-ended. Narrow tools (“summarize this matter’s deadlines”) outperform chat that pretends to practice law.

Document Intelligence

Hypothesis: Separating OCR text from field intelligence lets parsers survive DoI layout changes without retraining a whole model.
Prototype: Stage pipeline already in production intake — experimenting with per-issue layout fingerprints and confidence thresholds.
Technology: R2 · layout OCR · Workers field mappers · human review
Learning: OCR alone is never enough. Classification and review gates are the product; extraction is infrastructure.

Search Systems

Hypothesis: For 130K+ trademarks, prioritizing normalized exact and FTS tokens beats fuzzy “smart” matching when desks act on the result.
Prototype: Production search paths in NepalIPMS; lab work on controlled variant expansion after exact miss.
Technology: D1 FTS5 · KV hot cache · query planners
Learning: False conflicts destroy trust faster than a missed suggestion. Accuracy first; flexibility second.

Automation

Hypothesis: Renewal reminders only work when they attach to a verified matter record and a named owner — not a generic notification feed.
Prototype: Cron-triggered deadline checks tied to matter ownership; exploring smarter escalation when no one acknowledges.
Technology: Workers cron · D1 · notification hooks
Learning: Automation without ownership is noise. The hard part is the social graph of who must act.

Knowledge Graphs

Hypothesis: Explicit graph edges between companies, matters, and assets make portfolio questions (“everything for this client”) faster than nested joins alone.
Prototype: Domain model already encodes the hierarchy; exploring lightweight edge tables for related-party and opposition links.
Technology: D1 relational model · optional edge index · Workers API
Learning: Start from desk nouns. A graph that doesn’t match how lawyers talk about files becomes another abstraction to fight.

Human + AI Workflows

Hypothesis: Showing field-level confidence next to source pages makes practitioners faster at correcting AI than hiding uncertainty.
Prototype: Review queue patterns from OCR intake; testing denser “source + guess + confirm” layouts.
Technology: Portal UI · confidence scores · provenance IDs
Learning: Human+AI is a product surface, not a model choice. The review screen is where trust is won or lost.