🤓 Latest Submissions
.png&w=828&q=75)
ROADREAD: Precision OCR on AMD ROCm
ROADREAD is a high-precision, low-latency road sign and license plate recognition system engineered specifically for AMD ROCm hardware (AMD Radeon Pro W7900D, 48 GiB VRAM, RDNA3 gfx1100). Built for the AMD AI Developer Program Mini-Challenge 2, the pipeline accurately transcribes unseen and severely degraded imagery across six challenging domains: US license plates, Chinese license plates, word-based traffic signs, worded speed limits, advisory plaques, and multiline warning signs. To overcome the cold-start overhead of heavy machine learning frameworks on ROCm and strictly honor the 30-second per-image evaluation deadline, ROADREAD implements a decoupled client-worker architecture. A zero-dependency, standard-library Python client handles evaluator requests in under 10 ms, communicating over a local IPC socket with a pre-warmed resident worker daemon. The vision-language model (Qwen-VL in BF16 precision) remains resident in GPU memory, eliminating redundant weight loading and memory allocations across sequential requests. The transcription pipeline uses an adaptive escalation strategy: standard inputs are processed at a bounded 1.0 MP resolution, while ambiguous or severely degraded samples trigger a single high-resolution escalation pass under internal budget controls. Post-processing strictly enforces domain-specific formatting rules, including US banner and slogan scrubbing, full-width ASCII normalization, serial-only I/O digit correction for Chinese plates, and character-repetition safety without heuristic truncation. Validated under the official Docker container baseline (rocm/pytorch:rocm10.0_ubuntu26.04_py3.14_pytorch_release_2.13.0) within the 60 GiB uncompressed boundary, ROADREAD achieves 100% exact-match accuracy on published samples with a median latency of 0.565 seconds, and 90.8% exact-match accuracy on synthetic stress-test benchmarks with zero timeout violations.
1 Dec 2026
Say Less: Intelligent Voice Agent Repair Engine
THE PROBLEM: THE SENTENCE RESET You call your dentist from a noisy street: "Can I book a haircut on Thursday at 2 PM?" The voice bot mishears one word. It says: "Sorry, I didn't catch that. Could you repeat that?" Now you have to re-say all 11 words. The same street noise garbles the same word again. Standard voice agents discard the 10 words they heard correctly and reset the entire sentence. In call centers, this repetition loop drives over 40% of caller drop-offs. When a speech model mishears "Tuesday" as "Chewsday" or writes "9 AM" instead of "nine am", it often does so with 99% confidence. Blind confidence scores cannot catch the error, so the bot quietly writes the wrong appointment into the database. THE SOLUTION: SAY LESS Say Less asks about the specific word it missed rather than resetting the whole sentence. When a turn ends, Say Less checks whether the heard value belongs to that field's allowed catalog (7 weekdays, 10 services, 20 surnames). When a word lands out of catalog, Say Less ranks the closest valid matches by sound using Double Metaphone phonetic scoring. Then it asks the smallest question possible: 1. Name one candidate: "Tuesday?" 2. Offer two choices: "Tuesday or Thursday?" 3. Ask the category: "Which day?" 4. Transfer to a human. Every candidate offered to the caller is guaranteed to be a real, bookable value from the business schema. THE BENCHMARK We benchmarked Say Less on 26 real human recordings across mobile phones and the public SLURP corpus: Commit accuracy: 96% baseline jumped to 100% with Say Less. Words re-said per booking: cut from 0.77 down to 0.04 (a 95% reduction). Human escalations: dropped from 4% down to 0%. On acoustic stress tests, Say Less slashed caller repetition from 5.09 re-said words down to 0.07. At standard call center costs ($7 to $12 per cleaned-up booking), that saves $4,700 to $8,000 per 10,000 calls.
30 Sep 2026
