🤓 Latest Submissions
.png&w=828&q=75)
Pinpoint: exact web answers with their source
Pinpoint is a web retrieval agent built for AMD AI League Match 4. You ask for one specific piece of information in plain English, optionally with a page to start from, and it returns the value itself plus the URL of the page it was read from. When none of the pages it opened contains the value, the answer stays empty. The agent searches with keyless providers (DuckDuckGo, Bing, the Wikipedia API) or starts from the given URL, opens the pages and the links whose anchor text matches the request, and turns each page into clean text: navigation and citation footers removed, table rows kept together, JSON API responses flattened to field: value lines. The best passages go to Qwen2.5-7B-Instruct, which stays resident on the AMD GPU and replies with JSON: the value, the passages it used, an exact quote and an optional link to open next. Code then checks that the value really occurs in the cited passage. If it does not, the answer is dropped and the agent follows the suggested link, up to three read rounds inside a 24-second budget. Measured on an AMD Radeon Pro W7900D in the League notebook pool, inside the mandated ROCm 10 / PyTorch 2.13 / Python 3.14 image: 11 of 15 questions correct, 7.9 s mean and 16.1 s max per request in a fresh process, 15 GiB VRAM peak. The container follows the League contract (python3 /app/app.py, one JSON file per question in /app/output) and adds a single layer on top of the unmodified base layers. MIT licensed.
1 Dec 2026
Clearsheet: discharge sheets patients can follow
Patients leave hospital with a discharge summary written for clinicians: "Furosemide 40 mg PO BID (INCREASED from 20 mg daily)", "hold if heart rate below 55". Many of them read English as a second language. A language model can rewrite and translate that page in seconds, and it can also change a dose, drop a "stop taking", or swap a phone number, and nobody at the bedside can proofread a sheet in Ukrainian or Arabic. Clearsheet keeps the model in the drafting seat. Identifiers are masked on the hospital side first. Gemma 3 27B (bf16, one AMD Instinct MI300X, vLLM on ROCm) extracts medicines, follow-ups and warning signs, each with a verbatim source quote, then writes a plain-English line, the patient's language, and a literal back-translation. Deterministic code checks that every quote exists in the document, that every number sits inside its quote with its unit, that translations keep the dose, drug name and emergency action, and sends anticoagulants, insulin, opioids, stops and dose changes to a nurse. The nurse review console shows each line next to a highlighter mark on its source sentence; nothing prints until every flagged line has a decision. The printed sheet puts the patient's language first, English underneath, medicines as a pill-box grid. On 8 synthetic cases in 6 languages: 97% of medicines found, 100% exact doses, 0 invented medicines, 99% of planted wrong numbers blocked. One real catch: Gemma translated "call 911" as "call 103", Ukraine's ambulance number. Batched on one MI300X: 1118 output tokens/s.
18 Oct 2026
