Every invoice a company pays is a chance for delay, error or fraud. Processing a single invoice costs about $9.84 and takes 8.2 days on average, according to Ardent Partners in 2025. The FBI recorded more than $3 billion in losses from business email compromise in one year. Companies now want AI agents to handle this work, but an agent that reads invoices from outside senders can be fooled by one hidden line telling it to ignore its instructions and approve the payment. Flowforge is a working procure-to-pay platform built to solve that problem. Five AI agents named Intake, Guard, Risk, Policy and AI Review move each request from raw text to a decision. The principle is simple and strict. AI reads and suggests, fixed rules decide, and people approve the large payments. Three things set it apart. First, the approval rules are formally verified. Policy Studio tests every rule set against 240 possible situations and 7 safety properties in about one millisecond, shows an example for every failure, and lets an AI repair be checked again. Second, the Guard agent flags and rejects known injection attempts before any model sees them, and the rules rather than the model make the final call. Third, every action is saved in an audit log where each entry is chained to the one before it, so any edit can be detected. The platform runs on Vultr with two Cloud Compute servers behind a Load Balancer, Managed Postgres for shared data and Serverless Inference for the AI steps. It includes a dashboard, approvals, vendor records, Policy Studio, the audit ledger and a copilot. More servers can be added without changing how data is stored. We report results honestly. On 92 synthetic invoices, including 48 attack attempts, both a standalone AI approver and Flowforge approved zero unsafe invoices, so we do not claim better accuracy. The advantage is proof, speed, availability and auditability. Flowforge makes its decision in 0.017 milliseconds compared with about 2 seconds for the model.
Category tags:VeriFind AI is an autonomous AI research agent built for AMD AI League Match 4: Intelligent Web Information Retrieval. It accepts natural-language questions, navigates the live web, extracts structured evidence from multiple sources, and returns verified, citation-backed answers — refusing to fabricate any information it cannot prove. The Problem Web search returns lists of links. Users still have to read, cross-reference, and manually verify every source. VeriFind AI automates the entire research pipeline: from query understanding to evidence extraction to fact verification, delivering a single trusted answer with full source transparency. How It Works 1. Query Understanding — The user's natural-language question is parsed to identify the core entity, requested fields, and optimal search keywords using Google Gemini (gemini-2.0-flash). 2. Web Navigation — A serverless-aware browser engine (Playwright locally, httpx in production) crawls up to 8 pages across 2 depth levels, handling JavaScript-rendered content and respecting robots/timeouts. 3. Content Extraction — BeautifulSoup4 and Trafilatura extract clean text, headings, tables, and relevant links from each visited page. 4. AI Synthesis — Gemini is prompted with strict anti-hallucination instructions. It is explicitly forbidden from treating web content as trusted system instructions, eliminating prompt-injection risk. 5. Evidence Verification — Every claim in the answer is located verbatim in the retrieved page content. Citations that cannot be matched to actual retrieved text are rejected and the answer is set to null. 6. Conflict Detection — When multiple pages contain contradictory information, conflicts are flagged in the structured JSON response rather than silently merged.
JN
WebScout is an AI agent that takes a plain-English request, such as “find the return policy on this store’s site” or “what is the cheapest plan listed”, then navigates the website on its own and returns only the specific information asked for, along with the page it came from. Instead of dumping whole pages into a model, WebScout works in a loop. It plans what it needs, fetches a page, cleans the HTML into compact text and a list of links, decides whether the answer is present, and if not, follows the most promising link. A step limit and a visited-page list prevent endless loops, and every answer is checked against the source text, so the agent reports “not found” rather than guessing. The language model runs on AMD GPUs through ROCm in the AMD hackathon notebook environment, using an open-weight model served with vLLM to keep inference fast and cheap per page visited. WebScout is built for accuracy: it returns a short exact answer, the supporting snippet and the URL, so users can verify the result in seconds.
ROCm Raiders
A containerized OCR pipeline that reads US licence plates, Chinese plates and road signs from images, built on the mandated ROCm PyTorch base and run on AMD GPUs. Model: Qwen3-VL-8B-Instruct in bf16 with Hugging Face Transformers, greedy decoding. One prompt encodes the transcription rules: return only the registration number and drop state names and slogans around it, keep the Chinese province character and letter, copy characters exactly, and add no units or labels. Code post-processing joins multi-line text with single spaces and, for Chinese plates only, corrects O to 0 and I to 1 in the serial, which never uses those letters. Architecture: the harness starts a new process per image, and loading an 8B model takes longer than the 30-second limit. So the container's CMD starts a small server that loads the model once and warms up, and app.py is a thin client with a 25-second budget that always writes valid JSON, falling back to an empty answer if anything fails. Testing: beyond the 10 samples, I built two harder test sets: the samples under 10 degradations (blur, darkness, noise, skew, low resolution and combinations) and 46 synthetic plates and signs with unseen texts, also degraded. Results: 200/200 on the samples, 99/100 and 454/460 on the hard sets. Preprocessing with autocontrast was tried and dropped because it hurt accuracy. On an AMD Instinct MI300X: healthy after 23 s, 0.2-0.4 s per image, peak VRAM 18 GiB, image 45.7 GiB uncompressed. Built with AI assistance (Claude) for code review and test tooling.
Abrik-Ai
Ghostfox Zero is a two-match retrieval portfolio built entirely on AMD ROCm — one resident Qwen3-4B-Instruct server in bf16 on the AMD GPU, one BM25-first evidence pipeline, two containerized task front ends. MATCH 4 (Intelligent Web Information Retrieval): the agent navigates a saved copy of a website like a browser — every HTML page parsed, all navigation links followed, full BM25 index built over cleaned text, zero downloads at evaluation time. Given a natural-language query it extracts the exact value requested (name, date, price, version, phone number) with the source page as citation. MATCH 3 (Retrieval-Augmented Generation): the same fail-cited pipeline extended to mixed-document corpora — PDF (pypdf), DOCX and XLSX (stdlib zip+xml), CSV, source code, logs and config files all indexed with the same BM25 core, returning the short answer plus the exact files it came from, and nothing when the corpus does not contain the answer. Both containers implement the full league contract: phase 1 indexes the corpus and starts the resident model server within the startup budget, phase 2 answers each query as a separate process within the time limit, writing output JSON with answer, citations and confidence. Fail-closed by design: exact values only, citations always, empty answer instead of a guess — evidence-first where naive LLM agents hallucinate.
autokeren-solo
Maximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCode Maximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing
not sure
VeriFind AI is an autonomous AI research agent built for AMD AI League Match 4: Intelligent Web Information Retrieval. It accepts natural-language questions, navigates the live web, extracts structured evidence from multiple sources, and returns verified, citation-backed answers — refusing to fabricate any information it cannot prove. The Problem Web search returns lists of links. Users still have to read, cross-reference, and manually verify every source. VeriFind AI automates the entire research pipeline: from query understanding to evidence extraction to fact verification, delivering a single trusted answer with full source transparency. How It Works 1. Query Understanding — The user's natural-language question is parsed to identify the core entity, requested fields, and optimal search keywords using Google Gemini (gemini-2.0-flash). 2. Web Navigation — A serverless-aware browser engine (Playwright locally, httpx in production) crawls up to 8 pages across 2 depth levels, handling JavaScript-rendered content and respecting robots/timeouts. 3. Content Extraction — BeautifulSoup4 and Trafilatura extract clean text, headings, tables, and relevant links from each visited page. 4. AI Synthesis — Gemini is prompted with strict anti-hallucination instructions. It is explicitly forbidden from treating web content as trusted system instructions, eliminating prompt-injection risk. 5. Evidence Verification — Every claim in the answer is located verbatim in the retrieved page content. Citations that cannot be matched to actual retrieved text are rejected and the answer is set to null. 6. Conflict Detection — When multiple pages contain contradictory information, conflicts are flagged in the structured JSON response rather than silently merged.
JN
WebScout is an AI agent that takes a plain-English request, such as “find the return policy on this store’s site” or “what is the cheapest plan listed”, then navigates the website on its own and returns only the specific information asked for, along with the page it came from. Instead of dumping whole pages into a model, WebScout works in a loop. It plans what it needs, fetches a page, cleans the HTML into compact text and a list of links, decides whether the answer is present, and if not, follows the most promising link. A step limit and a visited-page list prevent endless loops, and every answer is checked against the source text, so the agent reports “not found” rather than guessing. The language model runs on AMD GPUs through ROCm in the AMD hackathon notebook environment, using an open-weight model served with vLLM to keep inference fast and cheap per page visited. WebScout is built for accuracy: it returns a short exact answer, the supporting snippet and the URL, so users can verify the result in seconds.
ROCm Raiders
A containerized OCR pipeline that reads US licence plates, Chinese plates and road signs from images, built on the mandated ROCm PyTorch base and run on AMD GPUs. Model: Qwen3-VL-8B-Instruct in bf16 with Hugging Face Transformers, greedy decoding. One prompt encodes the transcription rules: return only the registration number and drop state names and slogans around it, keep the Chinese province character and letter, copy characters exactly, and add no units or labels. Code post-processing joins multi-line text with single spaces and, for Chinese plates only, corrects O to 0 and I to 1 in the serial, which never uses those letters. Architecture: the harness starts a new process per image, and loading an 8B model takes longer than the 30-second limit. So the container's CMD starts a small server that loads the model once and warms up, and app.py is a thin client with a 25-second budget that always writes valid JSON, falling back to an empty answer if anything fails. Testing: beyond the 10 samples, I built two harder test sets: the samples under 10 degradations (blur, darkness, noise, skew, low resolution and combinations) and 46 synthetic plates and signs with unseen texts, also degraded. Results: 200/200 on the samples, 99/100 and 454/460 on the hard sets. Preprocessing with autocontrast was tried and dropped because it hurt accuracy. On an AMD Instinct MI300X: healthy after 23 s, 0.2-0.4 s per image, peak VRAM 18 GiB, image 45.7 GiB uncompressed. Built with AI assistance (Claude) for code review and test tooling.
Abrik-Ai
Ghostfox Zero is a two-match retrieval portfolio built entirely on AMD ROCm — one resident Qwen3-4B-Instruct server in bf16 on the AMD GPU, one BM25-first evidence pipeline, two containerized task front ends. MATCH 4 (Intelligent Web Information Retrieval): the agent navigates a saved copy of a website like a browser — every HTML page parsed, all navigation links followed, full BM25 index built over cleaned text, zero downloads at evaluation time. Given a natural-language query it extracts the exact value requested (name, date, price, version, phone number) with the source page as citation. MATCH 3 (Retrieval-Augmented Generation): the same fail-cited pipeline extended to mixed-document corpora — PDF (pypdf), DOCX and XLSX (stdlib zip+xml), CSV, source code, logs and config files all indexed with the same BM25 core, returning the short answer plus the exact files it came from, and nothing when the corpus does not contain the answer. Both containers implement the full league contract: phase 1 indexes the corpus and starts the resident model server within the startup budget, phase 2 answers each query as a separate process within the time limit, writing output JSON with answer, citations and confidence. Fail-closed by design: exact values only, citations always, empty answer instead of a guess — evidence-first where naive LLM agents hallucinate.
autokeren-solo
Maximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCode Maximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing credit only Only when I open OpenCodeMaximum power, size flexible Is the $50 a credit balance inside Render $50 existing
not sure