Skip to content
Chimeraforge

Systems I Built, Running Live

Working demos of systems I built: environments that score decision-making programs, checks on software agents and automated decisions, search, scheduling and rule engines, and a shared whiteboard. Each runs in your browser on made-up sample data and says where its numbers come from. Most end the same way: something that looks like success until a stricter check reads what actually happened. Built by Sahil Kadadekar.

  • 11 projects
  • 4 collections
  • Runs in your browser

Projects by collection

Flight Routing Lab

With a tight deadline, planning ahead gets 40 of 64 simulated trips to JFK on time where booking the earliest nonstop gets 0, but 8 fewer arrive at all.

Reinforcement learning ·Sample data

Offline Policy Evaluation Lab

A new decision policy looks 0.27 better on records of past decisions, but a control that ignores the situation gets 65% of that, and a changed reward erases it.

Reinforcement learning ·Sample data

Customer Service Environment

Two scripted runs refund the same $48.00 on a double-charged order: refunding the duplicate charge scores 1.00, refunding the real one scores −0.40.

Reinforcement learning ·Sample data

Code Verification Lab

Four candidate bug fixes: a two-test smoke suite passes three, the full six-test suite passes one. Passing the bug’s own example is not a fix.

Reinforcement learning ·Sample data

Spreadsheet Reasoning Inspector

Rules vote on which spreadsheet values are final: balanced voting marks a scratch cell final, and a stricter gate finds neither real result.

Agents & evaluation ·Sample data

Browser Agent Completion Gate

Six attempts at one booking: the success notice claims three, my earlier agent’s check accepts five, and only one saved the booking asked for.

Agents & evaluation ·Sample data

Intake Triage Scorer

All 12,672 signal combinations scored: shouting “URGENT” never moves a message’s priority, and once a safety gate fires, nothing else does.

Agents & evaluation ·Sample data

Staged Document Search

When a filtered search finds too few notes it drops filters: at its default it drops all three, says “ready”, and 2 of its 4 notes break the request.

Systems ·Sample data

Send Pacing Scheduler

In all 1,000 runs of its own example, the scheduler sends its last two messages before a person could have typed them.

Systems ·Sample data

Drone Mission Governance

With no telemetry at all the drone flies all 6 waypoints, and the mission’s lost-link instruction is never read; stale data mid-flight is only logged.

Systems ·Sample data

Collaborative Whiteboard

Two people recolour one note at once: in 6 of the 8 possible orderings a screen shows a colour the database lacks; numbered order fixes all 8.

Product ·Sample data