Mini projects
One build per module. Start with running code, add the missing capabilities, and prove the result.
Contract-first ingestion runner
Build an ingestion tool that receives fixture responses, validates each record, preserves failures, and emits a reproducible run envelope that a frontend could consume.
Escalation threshold decision report
Build an offline evaluation report that chooses an escalation threshold on validation examples, reports error costs and workload, and evaluates a frozen choice on a separate test fixture.
A tiny network you can audit
Build a small neural-network trainer whose forward pass, derivatives, and loss trajectory can all be inspected without a machine-learning framework.
Transformer resource worksheet
Build a reproducible resource report that compares cache layouts, prompt/output reservations, and idealized concurrency under an explicit memory budget.
Replayable model adapter
Build a model-integration adapter with an offline replay mode, explicit tool execution, strict result validation, and observable completion states.
Catalog agent workbench
Build a small local interface for a bounded catalog assistant, with a shared controller usable from a CLI and HTTP API and a visible evidence-backed run result.
Reservation gateway with an audit trail
Build a local reservation gateway that separates request envelopes, validation, authorization, and domain effects. Deliver a runnable CLI simulation plus documented production boundaries.
MCP boundary inspector
Build an inspector that explains why recorded tool attempts were accepted, rejected, or left uncertain, then connect it to one verified MCP server revision.
Evidence-led incident planner
Build a local incident planner that creates observable actions, consumes fixture observations, revises remaining work, and verifies the final diagnosis.
Inspectable incident control loop
Build a small controller with typed observations, working state, memory lookup, strategy choice, execution simulation, and a decision record for every action.
Memory lifecycle workbench
Build a local memory workbench with typed records, scoped retrieval, corrections, deletion, expiry, and a provenance view showing which summaries depend on which records.
Grounded configuration answer service
Build a local retrieval service over versioned configuration notes, with hybrid candidate fusion, explicit evidence bundles, and claim-level support checks.
Temporal dependency evidence explorer
Build a reproducible evidence explorer that answers which teams were connected to a deprecated dependency at a specified time, then evaluates passage and graph retrieval on labeled questions.
Recoverable release review coordinator
Build a small coordinator that executes independent review tasks, persists task outputs, and recovers without duplicating a final external action. The seed is an in-memory recovery simulation.
Inspectable invoice browser workflow
Build a browser workflow harness that records observations, proposals, preconditions, and persisted outcomes. The seed simulates a tiny invoice application without a browser or network.
Multimodal command evidence ledger
Build a timeline that joins spoken commands to screen observations and emits reviewable action proposals, retaining uncertainty when evidence is stale or missing.
Channel inbox and response router
Build a normalized inbox that admits one job per namespaced event and prepares channel-specific response drafts. The seed uses local envelopes and sends no messages.
Versioned collaborator gateway
Create a local gateway that selects an approved compatible collaborator, checks tenant and scope policy, and records task progress and accepted artifacts. The seed is a protocol teaching simulation.
Review surface with synchronized proposal state
Build a constrained generated review surface and a state reducer that rejects stale approvals. The seed combines local UI descriptions with local event state and prints a review preview.
Proactive incident proposal engine
Build an event-driven monitor that deduplicates failures, groups them into evidence windows, and creates versioned review proposals. The seed prepares proposals and performs no external notification.
A booking recovery journal
Build a reproducible simulator for reservations, payment capture, ticketing, and compensation. Produce an event journal and a crash-point recovery matrix that explain every externally visible effect.
A capability broker with an approval inbox
Build a local action broker for a synthetic report assistant. Its artifact is a policy decision journal and a review screen or command that shows the exact action before approval.
A routing replay and economics report
Build an offline replay tool comparing routing policies on identical tasks. Produce a report of task success, spending, escalation, and latency distribution with every synthetic assumption visible.
A reproducible agent change review
Build an evaluation runner and report for an original task dataset. The deliverable must connect a proposed agent change to paired outcome evidence, cost, uncertainty, and documented limitations.
A patch evidence workbench
Build a local workbench that indexes Python source, proposes impacted tests, and assembles a review packet containing the reproduction, patch, instructions used, and verification results.
A recoverable report-job service
Build a service that accepts report jobs, executes them in separate workers, checkpoints progress, and survives documented failures without duplicate publication. Deliver code, deployment instructions, and a failure-drill report.
Adaptation readiness report
Create a preflight report that combines dataset lineage, retained task coverage, protected-group checks, and adapter memory estimates before a training job is approved.
Reward shortcut audit bench
Build a small evaluation bench that exposes candidates which score highly under a proxy reward while failing an independent task criterion.
Serving capacity notebook without a GPU
Produce a reproducible capacity report that combines architecture-level cache math, page reservations, and scenario comparisons, then defines a future measured load test.
Tenant isolation regression pack
Create a synthetic regression pack and evidence report for a multitenant agent boundary, including cross-tenant access, scope errors, expiry, and benign utility.
A reproducible experiment dossier
Build a compact research artifact that states a falsifiable hypothesis, compares paired outcomes, records lineage, and supports a bounded conclusion.
Evidence desk: an integrated agent capstone
Deliver a tenant-aware incident evidence desk with a usable task interface, an MCP capability service, an A2A specialist integration, evaluation, operational controls, and a truthful repository report.
Seeds demonstrate a working core. Completing a project means implementing its milestones and acceptance checks, not just downloading the seed.