Team Drishtikon

Tracera turns a bug signal into a reviewed, verified fix, and never claims more than it observed. THE PROBLEM A bug arrives as a symptom: a sentence in a call, a screenshot, a screen recording. Turning it into a fix means finding the code, reproducing the failure, writing a change and proving it works. Teams do this by hand, and AI coding tools shortcut it with confident patches nobody can tell from a guess: the output says "fixed" without showing what was observed. WHAT TRACERA DOES Show Tracera a bug (a report, a screenshot, a screen recording, a Google Meet transcript, or a GitHub repository) and it: 1. Indexes the repository into a versioned code graph and answers questions with citations. 2. Reads screenshots and screen-recording keyframes with Meta's vision model, locates the offending element with SAM 3.1, and keeps what is reported, inferred and observed strictly apart. 3. Proposes cited hypotheses, then reproduces the bug in a locked-down Docker sandbox (no network, read-only filesystem, non-root, resource limits). 4. Generates a patch and verifies that exact patch against the original reproduction. The outcome (verified, partially verified, failed, blocked) is set by code rules from what the sandbox observed, never by a model. 5. Shows a person the diff, impact, evidence and a resolution report. Approval changes no files; applying, committing, pushing and opening a pull request are separate, explicitly authorized actions. WHAT WE BUILT AND RAN A real Google Meet call (a 53-second transcript) was imported; the model proposed four bug candidates with verbatim quotes that code verified; one was confirmed; Tracera ran a 10-stage investigation. It found the off-by-one in a free-shipping check, reproduced it in the sandbox with before/after screenshots, verified the one-line patch, and after approval applied it. The corrected page shows free shipping at $50.00. The same flow works from a screen recording and from a repository synced through a GitHub App. Around the core: a React web workspace, a Chrome extension that files a bug with redacted evidence, a stdio MCP server (14 tools) so coding agents can use Tracera, Google sign-in, and a hosted demo (Vercel + Render) next to a full local install with the Docker sandbox. 27 architecture diagrams and a narrated walkthrough are in the repository. HOW IT STAYS HONEST Models return strict-schema JSON that code validates; citations are checked; hypothesis states and verification outcomes are set only by rules; every result is labelled [live] or [mock]; repository files, logs, transcripts and screenshots are treated as untrusted data; secrets are redacted before every model call. Try it: https://tracera-gules.vercel.app (judge login: judge@tracera.example / Tracera-Judge-4271). Source: https://github.com/vishavmehra/Tracera Describe your use of Meta products (2,345 characters)