Thor v2 — RAG-Free Fitness Intelligence

Thor v2 is a domain-expert fitness AI built on a single fine-tuned Qwen3-8B model trained on 7,118 carefully constructed instruction-response pairs spanning exercise science, nutrition, programming, injury screening, and population-specific guidance. Unlike RAG-based fitness apps that retrieve documents at query time, Thor v2 encodes knowledge directly into model weights during supervised fine-tuning on AMD MI300X hardware using ROCm. Evidence is referenced through compact citation keys — e.g. [CITE:NSCA_HYPERTROPHY_VOLUME] — that the model emits inline. A lightweight citation resolver validates these keys against a locked registry and surfaces the source document on demand. If the model emits an unknown key, it is rejected at runtime. Hallucinated citations are structurally impossible. The dataset covers 113 unique citation keys from 9 authoritative organisations — NSCA, ACSM, ISSN, NASM, HHS, USDA, NIH, CDC, and ExRx — with 80 exercise technique entries and 14 population profiles including senior, postpartum, teen, vegan, rehab return, and competitive athlete. Six conversational style variants (casual, research-nerd, anxious, skeptical, verbose, follow-up-first) are baked into training so the model adapts tone naturally without prompt engineering. Training results: 100% JSON contract pass rate across all eval prompts. Coach gating behavior confirmed — model asks clarifying questions before prescribing when context is missing, rather than giving generic advice. All responses emit valid citation_keys, follow_up_questions, and safety_notes fields. Adapter size: <350MB on top of a frozen 8B base. Built entirely on AMD MI300X (192GB HBM3, ROCm 6.3) using HuggingFace PEFT + TRL. One model. No retrieval. No vector database. The model knows. The resolver proves.

Agronomist

Resume Compact

Resume Compact, the HR Resume Analyzer, is an intelligent, end-to-end recruitment automation platform designed to streamline candidate screening and evaluation by combining modern document processing, artificial intelligence, and web technologies to eliminate manual resume review bottlenecks and provide data-driven hiring insights to recruitment teams. Key technical features include seamless fallback PDF processing with transition from native text extraction to OCR when needed, LangChain structured output using StructuredPrompt and Pydantic to ensure JSON-compliant responses, strict schema enforcement that validates all outputs against Pydantic models and rejects malformed responses, detailed and context-rich prompt engineering to guide the LLM toward accurate structured analysis, async HTTP using httpx for efficient non-blocking calls to the LLM API, robust exception handling with informative HTTP status codes and error messages, and comprehensive audit logging with timestamps, file IDs, and LLM raw responses stored for compliance and debugging. The system reduces recruitment cycle time by 60-80%, eliminates subjective bias in initial screening, allows HR teams to focus on relationship-building rather than document review, and, by standardizing candidate data into structured JSON, enables downstream analytics, reporting, and integration with ATS (Applicant Tracking Systems).

ROCmPort AI

CUDA-to-ROCm migration fails not at translation but at correctness. hipify-clang renames APIs mechanically. It cannot detect that a reduction kernel assuming warpSize=32 will silently produce wrong results on AMD wavefront-64. Lanes 32 through 63 are skipped. The code compiles. The output is wrong. Nobody tells you. ROCmPort AI is a closed-loop agentic system built to catch exactly this class of bug before it reaches production. The pipeline runs five agents in sequence. The Analyzer performs a static scan before any LLM call, grounding the system in what is actually in the code. The Translator runs hipify-clang as a first pass then applies LLM corrections for architecture-specific issues the tool cannot handle. The Optimizer applies MI300X-specific changes: wavefront-64 alignment, LDS bank conflict padding, shared memory tiling. The Tester compiles with hipcc and profiles with rocprof on real AMD hardware. The Coordinator evaluates the profiler output and decides whether to iterate or finalize. All four demo kernels were compiled and profiled on a real AMD Instinct MI300X on AMD Developer Cloud. gfx942. ROCm 7.2. data_source: real_rocm. The primary model is Qwen2.5-Coder-32B-Instruct, purpose-built for code reasoning tasks. Groq LLaMA-3.3-70B handles log parsing as a cost-efficient fallback. In production, Qwen runs via vLLM on the MI300X instance itself. Failure cases are documented explicitly, including library-heavy CUDA using CUB, Thrust, or cuDNN, which requires manual review after ROCmPort output. This is intentional. Credibility requires honesty about scope. The value is not speed. It is correctness before execution, and a decision trace that a senior engineer can audit.

Raksha Setu: AI emergency bridge

In India and across the developing world, over 2 billion people lack reliable emergency medical response. The average ambulance response time is 18–20 minutes. Cardiac arrest causes irreversible brain damage in 4. That 14-minute gap is not a medical failure, it is an infrastructure gap. Nobody owns it. Raksha Setu does. Raksha Setu (Sanskrit: the bridge) is an AI-powered emergency response mesh built on AMD Developer Cloud. When a distress call is received, Qwen 2.5 running on AMD MI300X performs medical triage in under 2 seconds, classifying case severity, type, and the minimum responder qualification required. It then dispatches the three nearest trained Angel volunteers from a verified, gamified network ranked by certification level. The first to confirm arrives on site in under 5 minutes, guided step-by-step by real-time AI protocol streamed to their phone. When the ambulance eventually arrives, it receives a structured AI-generated handoff log, vitals, timeline, and all actions taken, before the crew even enters the building. AMD MI300X is not incidental to this product. Inference latency here is measured in lives. Every millisecond of triage delay is a millisecond closer to permanent brain damage. ROCm gives us the speed; the MI300X gives us the scale to handle concurrent emergencies across an entire city. Raksha Setu is built as infrastructure, not an app. The same Angel volunteer mesh powers Raksha Doot, our Phase 2 women's safety layer, using identical volunteers, identical AMD compute, and an AI threat classification system that closes a different but equally critical gap.

Hands for AI

Chrome offers an unusually deep programmatic surface to AI agents — the DevTools Protocol, the accessibility tree, JavaScript evaluation, network interception, multi-tab and multi-context isolation. For most of the web, that surface is enough to skip visual reasoning entirely from the first action, replacing screenshot-and-guess with structured targeting that survives DOM changes. This project drives all of it through a single MCP server, with accessibility references and a six-rung escalation ladder that auto-selects targeting strategy. Vision and OCR remain available as fallbacks for canvas apps, custom-drawn UIs, and anything without useful structure. What makes the architecture more than a Chrome controller is the workflow layer alongside it. While the agent works, network traffic is captured and the underlying API patterns extracted into replayable flows. On subsequent runs the agent skips the browser entirely and executes direct HTTP calls — millisecond execution at a fraction of the inference cost. Credentials and TOTP seeds live in an OS-level vault, so replay works across sessions and machines without secrets ever touching chat context. Optional personal-assistant mode extends the same engine beyond Chrome to native Windows applications via UI Automation and OCR-based control of anything else, for workflows where browser scope isn't enough. Most agentic browser tooling today is either tied to a vendor's cloud or wraps a single automation library. This stack stays fully open, MCP-native, and self-hostable — running over local STDIO, which removes the network hop between agent and tool at every step and meaningfully cuts step latency on multi-action tasks. Composable with any compatible host and any other MCP tool. The reasoning layer runs on AMD Developer Cloud with ROCm-hosted open vision and language models. No proprietary inference dependencies anywhere in the stack.