Sankar Subbayya@sankars

Enterprises deploying AI agents face three production blockers: no audit trail, no policy enforcement, no cost accountability. Agent Sentinel solves all three. Drop Sentinel in front of any MCP- or A2A-speaking agent. Every tool call passes through a four-stage pipeline: a sub-5ms static engine (regex denylists, role ACL, refund caps), a cheap drift detector for prompt-injection markers, an inline Gemini 2.5 Flash gate with structured output, and — for the ambiguous 20% — Gemini 2.5 Pro escalation that reads whole policy documents via Cached Content. No chunking, no RAG drift. Each decision becomes a hash-chained, HMAC-signed audit receipt that cites the exact policy version. The same loop emits a per-business-unit cost event for chargeback. Sentinel also publishes its own Google A2A agent card at /.well-known/agent.json so peer agents can discover the governance plane. Verified end-to-end with real Gemini: 96.8% accuracy across 155 labeled scenarios spanning 12 categories (happy path, PII variants, injection variants, role matrix, regulatory, edge cases, drift); 88 passing pytests; INTEGRITY: PASS on the hash-chained ledger after a tamper-injection test; 5,000 receipts load-tested at 800 req/s with zero chain breaks. Live at https://agent-sentinel.up.railway.app. Three buyers, one architecture: compliance officers get queryable evidence with cited policy versions, CISOs get inline enforcement plus prompt-injection defense, CFOs get per-BU cost attribution. Built end-to-end on Google's stack — Gemini 2.5 Flash + Pro, Cached Content, Files API, A2A, and a flagship adapter for Google ADK. Cloudflare for AI agents.

Path to Care is a multimodal, agentic decision-support system for rural healthcare — built in 24 hours on one AMD Instinct MI300X. A patient 18 km from the clinic, earning ₹350/day, faces ₹180 transport, an active harvest, and a swelling foot. Should he go? The system never tells him what's wrong — it helps him decide. **Cardinal rule (enforced in code).** Always "signs suggest", never "you have." Image output is always **top-3 with confidence** — single-class is impossible by construction. A regex post-filter rewrites diagnostic phrasing on every output, logged for audit. **What it does.** (1) Ranks skin conditions as top-3 + confidence; (2) assigns Red/Yellow/Green urgency; (3) flags red signs; (4) frames cost-benefit using local barriers (distance, wage %, harvest, drug-stock). Output: patient brief + pre-visit SOAP. **Architecture.** Five MCPs orchestrated by a DSPy-style ReAct loop: `camera_capture`, `image_classifier`, `soap_extractor` (Qwen-2.5-7B), `village_context`, `triage_reasoner`. Image-classifier and triage-reasoner share one loaded **Gemma 4 31B-it**. **Serving.** MI300X 192 GB, ROCm 6.3. `vllm/vllm-openai-rocm:v0.20.1` Docker, base + LoRA served simultaneously via `--enable-lora`. HF Space (Next.js/React) proxies every triage request to the droplet. **Fine-tuning.** LoRA on Google's SCIN dermatology dataset (top-16 conditions, top-k probability targets): **top-1 28.0% → 35.0% (+7.0 pp)** on a 100-case holdout. First positive delta after 3 documented failures — lesson: per-class samples > epochs. MedGemma 27B baseline scored 3 pp worse than Gemma 4 31B. **Frontend.** Patient / Clinician / Audit tabs · browser camera · Web Speech voice dictation · audit panel showing every MCP tool invocation. **v2.** Village fieldwork · 80-case set · physician review · Fitzpatrick-stratified bias audit · Tamil UX · GRPO/RLVR. 🤗 sankara68-path-to-care-react.hf.space · 💻 github.com/SankarSubbayya/path_to_care

Midstream solves a critical problem in AI: users pay upfront for output regardless of quality. Whether a research run drifts, code breaks, or an image is off-prompt, you've already paid full price. This wastes 30–70% of AI spending annually. Our solution: quality-gated streaming payments. The LLM generates output in 32-token chunks at $0.0005 each. Between chunks, the buyer's quality oracle (LLM, compiler, CLIP, etc.) assesses whether to continue paying. If quality drops below threshold, the buyer stops authorizing, the stream ends, and they pay only for the prefix that passed. We've built a working MVP on Circle Nanopayments + Arc testnet. Circle's batching achieves <1% overhead per chunk—impossible on Stripe ($0.30 minimum, 60,000% overhead) or Ethereum ($0.30 gas). Our demo shows: - Session 1: Full run, 31 chunks, $0.0155 - Session 2: Quality drop at chunk 13, killed, 60% saved - Session 3: Rapid drift at chunk 4, killed, 87% saved - Session 4: Compiler catches broken code at chunk 9, killed, 68% saved Total: 58 chunks paid, $0.0290 spent (vs. $0.0620 full run) = $0.0330 saved. We've verified 213 Circle Transfer UUIDs settled on Arc testnet via Circle Gateway batching. Every settlement is clickable on testnet.arcscan.app. The architecture is oracle-agnostic: works with any quality signal (text, code, image, video, schema validation, etc.). This is a platform, not a one-off.

Grab your ideas, spin up compute, and start building what’s next! Start with $100 in AMD Developer Cloud credits and begin building right away. Get access to hands-on tutorials, workshops, and curated learning resources to help you build faster and go further. To unlock credits and join the hackathon, sign up for the AMD AI Developer Program. 💰 $10,000+ total prize pool Developers, builders, and innovators - bring your boldest ideas and let’s turn them into real AI systems on AMD.

📅 May 11–19, 2026 • May 11 – 19: Online Build Phase. • May 18 (Hybrid Build Day – Online & Onsite) • May 19 (Demos & Awards – Online & Onsite) 🌍 Hybrid event. Join online from anywhere or build onsite at the venue. 🏆 Winners will be announced live on stage at the AI & Big Data Expo North America on May 19 📍 On-site Venue (May 18–19): San Jose McEnery Convention Center, CA, USA 🤝 Join solo or with a team. 🎟 FREE AI & Big Data Expo North America ticket included for participants that are able to join the on-site phase. 🏆Compete for a $10,000 prize pool 🧑‍💻 Apply now to design, prototype, and showcase enterprise-ready innovation.

Grab your ideas, spin up compute, and start building what’s next! Start with $100 in AMD Developer Cloud credits and begin building right away. Get access to hands-on tutorials, workshops, and curated learning resources to help you build faster and go further. To unlock credits and join the hackathon, sign up for the AMD AI Developer Program. 🤝 Go solo or team up — ship something real. 💰 $21,500+ total prize pool & 1 × AMD Radeon™ AI PRO R9700 GPU up for grabs. 📍 On-site Venue (May 9-10): MindsDB SF AI Collective 3154 17th St, San Francisco, California, USA 📍On-site participation is by invitation only. Travel and accommodation expenses will not be covered. 🧑‍💻 Developers, builders, and innovators - bring your boldest ideas and let’s turn them into real AI systems on AMD.

In this hybrid hackathon powered by Arc and Circle, developers will build the next generation of Agentic economic applications using Circle’s Nanopayments infrastructure with settlement on Arc in USDC. For the first time, high-frequency, usage-based transactions between users, APIs, and AI agents can operate without gas overhead eroding margin. Arc provides stablecoin-native settlement with predictable, dollar-denominated USDC fees and deterministic, sub-second finality — enabling real-time coordination and economically viable per-action pricing at scale. 📅 April 20 – 26, 2026 • Apr 20 – 25 (Online): Submit projects by Apr 25. • Apr 25 (On-site Build – SF, CA): Teams refine projects. • Apr 26 (On-site Demos & Awards – SF, CA): Live pitches & winners announced. Get guidance from Arc and Circle experts through on-site support. 📍 On-site Venue (April 25-26): MindsDB SF AI Collective 3154 17th St, San Francisco, California, USA 🤝 Go solo or team up, it's your flow. 🏆 Compete for a $15,000+ Prize Pool. 📍 Please note: If approved for the on-site experience, you’ll also have the opportunity to showcase your project live. Travel and accommodation will not be covered. 🧑🏻‍💻 Secure your spot now – sign up before the Kick-Off Stream!