CodePulse
Proctored technical interviews, and games that make them worth sitting
- Role
- Full-stack developer: architecture, proctoring engine, interface
- Year
- 2025
- Discipline
- Full Stack · Real-Time Systems · Product
- Context
- A web platform for conducting and scoring remote technical interviews, with an integrity layer and a set of timed assessment games.
Remote hiring broke the technical interview. The moment the candidate is alone with a browser, a second monitor and a language model, a take-home question stops measuring anything. The usual response is surveillance: lock the browser, watch the webcam, flag every glance away from the screen. It works, in the sense that it produces a number. It also makes good candidates refuse to interview.
Code Pulse takes the opposite position. Integrity comes from how the session is instrumented rather than from how aggressively the candidate is watched, and the assessment itself is built to be hard to fake. The scoring is not the final answer but the path to it: keystroke cadence, edit history, how someone recovers from a failing test. Copying a finished solution produces a suspiciously clean recording, and the recording is the evidence.
What madethis hard.
- 01
Proctoring signals are individually meaningless
Tab blur happens when a notification arrives. A paste happens when someone moves their own scratch code across. A long pause happens when someone is thinking. Any single signal used as a verdict generates false accusations against honest candidates, which is worse than missing a cheat.
- 02
The browser is a hostile environment for monitoring
Page visibility, focus and clipboard events are all advisory. They fire inconsistently across browsers, and anything running purely client-side can be disabled by whoever controls the client. Trusting the candidate's browser to report on the candidate is circular.
- 03
Running untrusted code safely
Candidates submit arbitrary code that has to compile and execute against test cases. Executed naively that is remote code execution as a product feature: infinite loops, fork bombs, filesystem access, outbound network calls.
- 04
Latency shows up as unfairness
In a live session, the interviewer's editor and the candidate's have to agree. Naive synchronisation drops characters when both type at once, and a candidate who loses two seconds of work under a timer has been penalised by the tool rather than by the question.
- 05
Games have to measure something real
A timed puzzle that rewards having seen the trick before measures recall, not engineering. The format has to stay resistant to memorisation while remaining solvable under pressure by someone competent.
How Isolved it.
- 01
Score integrity as a weighted profile, never a single flag
Signals are collected independently and combined into a session profile: focus changes, paste volume relative to typed volume, typing cadence variance, gaps between edit bursts. Nothing auto-rejects. The output is a ranked review queue for a human, with the specific evidence attached.
- 02
Keep the source of truth on the server
Edits stream to the backend as an append-only operation log rather than being summarised by the client. Timeline, replay and integrity scoring are all derived server-side, so disabling client instrumentation removes the candidate's data rather than hiding their behaviour.
- 03
Execute submissions in disposable containers
Each run gets its own short-lived container with no network, a read-only filesystem, and hard CPU, memory and wall-clock limits. The container is destroyed after the run, so one submission cannot affect the next.
- 04
Real-time editing over an ordered channel
The session runs over WebSockets with server-assigned sequencing, so concurrent edits converge to the same document instead of racing. Disconnects replay from the last acknowledged operation rather than losing the buffer.
- 05
Design the games around process, not trivia
Timed challenges built to have several viable routes: debug a failing suite, extend a system under a constraint, optimise something already correct. The replay shows how the candidate got there, which is the part that resists being pasted in.
What itdoes now.
Integrity reported as a reviewable profile with attached evidence, so a human makes the call rather than a threshold.
Server-authoritative edit history, which means the session replay survives a candidate tampering with their own client.
Sandboxed execution with enforced CPU, memory and time ceilings and no outbound network.
Live collaborative editing that converges under concurrent typing and resumes cleanly after a dropped connection.
An assessment format built around process rather than recall, with the working visible to the reviewer.
Frontend
- JavaScript
- HTML
- CSS
- WebSockets
Backend
- Python
- REST APIs
- Async Workers
Data
- PostgreSQL
- Event Log
- Schema Design
Runtime
- Docker
- Linux
- Sandboxed Execution
AlumniPortal
A role-based platform connecting graduates, students and faculty: directory, events, opportunities and mentorship, built on a normalised relational schema with security treated as a requirement rather than a patch.
Read case study