Working systems · Honest status
Proofs before platforms.
Each proof of concept tests one architectural claim under real operational constraints, rather than whether a technology can produce a polished demo.
The registry records what is running, what remains unverified, and what I learned from the boundary being tested.
Seven focused systems · verification status on each entry
Browse the registry-
Research lifecycle assistant for physics & math PhDs
Where in a PhD research loop does comparing multiple models actually change what gets published?
A single research question moves through six stages (literature retrieval, hypothesis and derivation, model comparison, numerical simulation, peer critique, and write-up), with several real models compared side by side at the stage where their answers diverge.
- Multi-model comparison
- Vanilla JS
- Mock research data
-
Fraud case investigation
Can a long-running investigation stay durable across a human review SLA without losing state if a worker restarts?
A Temporal workflow runs automated fraud checks, retries a flaky device-intelligence call without hand-written retry logic, then waits on an analyst's signal against an SLA timer. Killing and restarting the worker mid-case resumes exactly where it left off: nothing already completed re-runs, nothing in flight is lost.
- Temporal
- Python
- FastAPI
- PostgreSQL
-
Database column masking by role
Should the same query return a different record shape when the caller's responsibilities change?
This test moves role-scoped visibility to the database boundary. Support and finance users see different portions of the same customer record without relying on application code to hide sensitive columns.
- PostgreSQL
- Column security
- Role-based views
-
Scoped financial-data tools for agents
Can an agent receive only the financial-data capabilities its current task is allowed to use?
An MCP server exposes scoped tools instead of a standing API key. Every call is constrained by the calling context and produces an audit trail that can be reviewed independently of the agent conversation.
- Model Context Protocol
- Python
- Scoped authorization
-
Orders and payments streaming pipeline
Can event-time correlation remain correct through duplicates, late arrivals, and checkpoint recovery?
Kafka decouples ingestion while Flink joins payments to orders in event time, deduplicates by event ID, and recovers without double-counting. PostgreSQL serves current state and minute-level revenue metrics for live verification.
- Kafka
- Flink
- PostgreSQL
- Event-time processing
-
Clinical evidence review
Can AI-assisted screening reduce reviewer effort without weakening recall or traceability?
A Temporal workflow coordinates literature search, structured model screening, and human verification. A worker can fail mid-search and resume without repeating completed work. Model suggestions and reviewer decisions remain separate so corrections propagate by construction.
- Temporal
- FastAPI
- React
- PostgreSQL
- Structured outputs
-
Enterprise agent authorization gateway
Can a sub-agent ever obtain more authority than the parent agent that delegated the task?
A complete delegation path (user → agent → sub-agent → gateway → policy decision → MCP tool) tested end to end: identity propagation, delegated authorization, Cedar policy evaluation, short-lived capability issuance, MCP-level enforcement, a LangGraph execution path, and an immutable audit trail. Beyond the happy path, the design is tested against specific attacks: a prompt requesting an unauthorized tool, a sub-agent attempting privilege escalation, a tool called outside its task scope, an expired delegation still presented as valid, a confused-deputy scenario, a policy change mid-execution, and an action that requires human approval before it can proceed.
- Cedar policy engine
- LangGraph
- Model Context Protocol
- Delegated authorization