suvasis mukherjee@suvasism

**M3I — Targeted LoRA-DPO via Multi-Observer Attribution** When small LLMs fail at multi-step reasoning, the errors rarely spread uniformly — specific layers carry the failure. Standard LoRA is blind to this. M3I asks: can two LLM observers identify those layers, then train LoRA-DPO on only them? Five-phase pipeline, validated end-to-end on DeepSeek-V2-Lite (16B MoE): 1. Inference + per-layer signal capture via custom ggml hook 2. Cross-observer audit (Claude Sonnet 4.5 + Gemini 2.5 Pro), agreement required 3. Statistical attribution: paired t-tests at failure vs control tokens, Bonferroni-corrected 4. LoRA-DPO on attribution-identified layers only 5. Held-out eval: verdict + position-bias-controlled pairwise **Status (May 10, 2026):** Trained on 78 problems (50 PhD-level math + 28 crypto microstructure), 116 preference pairs, 3 epochs. Attribution flagged layers {6, 7, 8, 10, 13} — strongest signals at L13 attn (Z=+3.71), L10 attn (Z=+3.13), L6 MoE (Z=-2.84). Held-out eval on 12 problems: base 3 / LoRA 1 / tie 8 in pairwise wins. Three problems shift "incorrect" → "insufficient" — behavioral change without quality gain. Net: null to slightly negative. **Caveats:** n=12 held-out is small; ~20% adapter capacity stripped for GGUF conversion (kv_b_proj / MLA incompatibility); 116 pairs is modest; 16B is below capability floor for these problems. **Next:** (a) random-layer control to separate "attribution fails" from "any DPO at this scale fails"; (b) seed sensitivity for error bars. ~90 min MI300X compute + ~$2 API spend → defensible methodology paper either way. Inference: AMD MI300X via llama.cpp with HIP backend, ~115 t/s with adapter.

MEV Shield is a real-time pre-trade intelligence layer for DeFi traders and autonomous agents. It captures live Ethereum mempool transactions from multiple sources (PublicNode, Alchemy, Infura), scoring each pending swap using proprietary microstructure signals — Hawkes process order-flow clustering, Kyle's Lambda price impact, and VPIN toxicity — to detect sandwich attacks and frontrunning risk before execution. Each transaction receives a cliff score (VPIN × LOB depth × gas Z-score) that quantifies extraction probability. Pools are monitored continuously with per-pool intelligence cards showing Hawkes intensity, Kyle's Lambda impact, and next-attack timing estimates. Bot fingerprinting clusters known MEV addresses by behavioral signature. Intelligence is delivered as a metered API where consumers pay $0.01 per alert via Nitrolite ERC-7824 off-chain state channels, eliminating per-query gas costs. Payments are cryptographically signed ECDSA state updates — verifiable without touching the chain. When accumulated payments reach a threshold, the channel closes via a Flashbots MEV-protected transaction and settles autonomously as a real USDC transfer on-chain via Circle Arc developer-controlled wallets, confirmed on ETH-Sepolia. Claude, AIsa.one provides pre-trade risk assessments on every fifth alert, combining quantitative microstructure signals with natural language recommendations covering slippage tolerance, private mempool routing, and sandwich guard configuration. AI agents can autonomously consume and pay for real-time financial intelligence. No subscriptions, no API keys per user, no manual settlement — just a self-sustaining micropayment loop. Live mempool data flows in, off-chain ECDSA payments accumulate, on-chain USDC settles when the threshold is reached. DeFi traders lose an estimated $1-3B annually to MEV extraction. MEV Shield is the pre-trade defense layer that makes institutional-grade microstructure analysis accessible at $0.01 per query.

orgeClaw × Kraken is an autonomous crypto trading agent built on a production-grade stack that goes far beyond a simple trading bot. The agent pipeline executes 7 sequential Temporal activities: connecting to the Kraken CLI MCP server for market data, fetching AI signals from PrismaAPI, computing RSI(14) and VWAP deviation analysis with volume confirmation, gating each trade behind an ERC-8004 USDC micropayment for trustless execution, executing paper trades with 10% position limits and 2% stop loss enforcement, tracking realized PnL with a FINRA-style SQLite audit log, and delivering formatted trade summaries to Slack. ForgeClaw acts as the design-time layer — a BPMN agent composer (forgeclaw-app.vercel.app) that generates Temporal workflows from visual pipelines. VerifyClaw scans every agent skill against 25 SAFE-MCP threat patterns before deployment, with a max risk score of 3/100 on this agent. Redpanda streams all trade events across 5 Kafka-compatible topics in real time. The dashboard is a pixel-accurate Kraken Pro replica with live signal feed, agent workflow panel, executor, trade history, PnL analytics, open orders, Slack log, and ERC-8004 payment ledger — all backed by a FastAPI service proxied through nginx and pulling from SQLite on every cycle. Clicking Run Agent fires a real Temporal workflow end-to-end, not a simulation. Infrastructure: Temporal + PostgreSQL for durable orchestration, Redpanda for event streaming, nginx reverse proxy, Docker Compose for one-command deployment.

ChartSeek: AI-Powered Trading Education Video Intelligence Traders face an overwhelming challenge: thousands of hours of educational videos, yet finding that moment explaining a "head and shoulders pattern" means scrubbing through endless footage. Traditional search fails because traders need to find visual chart patterns, not just spoken words. The Industry Gap Current video platforms offer only keyword search against titles and descriptions. Trading platforms like TradingView, Investopedia, and YouTube provide no way to search inside video content. Enterprise solutions cost $50K+ annually yet still can't match visual patterns. Traders waste hours rewatching purchased content, unable to locate specific setups. This gap costs traders their most valuable resource: time for analyzing live markets. How ChartSeek Bridges This Gap ChartSeek combines OpenAI's CLIP visual understanding with Whisper speech recognition. Unlike keyword search, ChartSeek understands what's visually on screen. Search "bullish engulfing on support" or "descending triangle breakdown"—and instantly jump to that exact frame, even if never explicitly mentioned by the instructor. The system transcribes spoken commentary, extracts representative keyframes, and generates visual embeddings. Searches query both transcript and visual index simultaneously, returning ranked results with confidence scores. Technical Foundation Built on TheAgenticAI's CortexON multi-agent framework with OpenAI Codex workflow architecture. Runs 100% locally using open-source models—zero API costs, complete privacy for proprietary strategies. Key Capabilities - Visual pattern search by description - Cross-modal text-to-image matching - Automatic timestamped transcription - Instant clip extraction ChartSeek delivers 90% reduction in search time, transforming passive video libraries into queryable intelligence. Less searching, more trading.

Problem: Monthly subscriptions ($50-$500) for MEV analytics waste capacity, block AI agents from autonomous access, and traditional payment fees (~$0.30+2.9%) make pay-per-query impossible. Solution: x402 micropayments enable humans and AI agents to pay only for what they use. How it works: x402 Protocol: Returns HTTP 402 with exact price ($0.01-$0.05/query)—machine-readable pricing replaces opaque tiers Circle Wallets: Onchain treasury with guardrails ($1/tx, $10/day)—AI agents spend autonomously without human approval Arc Network: Near-zero fees (~$0.003) settle $0.05 payments in ~2 seconds using native USDC gas.

⏱️ Build the next generation of Autonomous Agents in the heart of Europe’s AI Revolution. 📅 May 13 – 20, 2026 May 13 – 19: Online Build Phase. May 19: An exclusive on-site build day at Milan AI Week for selected participants. May 20: Awards Ceremony. 📍On-site Venue (May 19–20): Fiera Milano (Rho), Milan, Italy 🤝 Join solo or with a team. 🎟 FREE Milan AI Week conference ticket included for participants. 💰 $32,000+ prize pool 🧑‍💻 Apply now to turn your autonomous agent into a market-ready enterprise solution.

📅 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.