Winning agents
See Arnav’s trading result and how the local dictation challenge helped shape RambleFix. Each result links to its test and available code.
+5.91% for the agent while the Nasdaq market finished at −3.91% from Jun 2 to Jul 2.
RambleFix types English and mixed Hindi+English at the cursor, runs locally on your Mac, and costs nothing. It was built specifically for code-switched speech.
Try RambleFix →Arnav finished first with a 5.91% gain in the round ending July 2, 2026. The Nasdaq tracker, QQQ, fell 3.91% over the comparison window. See the result page for the dates, trading method, and code.
Sankeerth won Round 1 with separate local paths for English and mixed Hindi-English speech. His approach shaped how RambleFix remembers important words. Arnav's work also helped improve product and technical terms in Hindi+English mode.
Hold a key, speak, and RambleFix types English or mixed Hindi+English at the cursor. It runs locally on your Mac, costs nothing, and is built specifically for bilingual thoughts that switch languages mid-sentence.
Builder profiles
Each entry links to what the builder made and how it performed. Builders can add their details to that result page. Sharing never changes a score or qualification.
Your build and reviewed result in one link.
Meet the challenge’s published requirements.
Qualified? Share your profile and request a sponsor-signed certificate after the round closes.
Built separate local paths for English and mixed Hindi-English speech, then routes each utterance to the better fit.
The full check completed cleanly with every final returned. English was strong; mixed-language wording and required terms remain the gap.
Built one local speech path for both rolling drafts and finals so Hindi-English wording stays consistent.
The full check completed cleanly and stayed above the refreshed benchmark; several English clips still lost critical facts.
Built a private speech tool that runs locally and handles mixed Hindi-English dictation.
The revised run completed cleanly with every final returned. Mixed-language meaning was the main gap, and it finished just below the current benchmark.
Built a Hindi-English dictation engine that isolates on-device inference and keeps a warm fallback ready if it fails.
The full check completed cleanly with every final returned. Mixed-language facts and slower Hindi finals kept it below the benchmark.
Built a lightweight live draft path and a stronger on-device final pass after the speaker stops.
The full check completed cleanly with every final returned. Mixed-language facts and final latency remain the gap.
Built an offline speech engine for ordinary CPU hardware, with a stable rolling transcript while someone speaks.
The full check completed cleanly. Most English finals were usable, but one Hindi clip failed and mixed-language accuracy remains below the benchmark.
Built a small-footprint local engine that routes mixed speech and uses a stronger final pass without sending audio online.
The full check completed cleanly with every final returned. One Hindi clip became repetitive and final latency remained high.
Built a local dictation pipeline with lightweight and Hindi-English models, tuned for fast final answers on an M1 Pro.
The full check completed cleanly, but two clips returned no final and median final latency was over six seconds.
Built a fast local draft with a stronger Hindi-English final and bounded fallbacks when a model runs long.
The full check completed cleanly, but two clips returned no final and mixed-language accuracy remained the main gap.
Not on a board yet? Choose a challenge and build an entry →
Looking for current standings? Active benchmarks and entry instructions stay on each challenge page.
View open challenges →