Mirrored from filpgf.io — ProPGF Batch 3 (Karma program 1479, application 6a3194ae13ef2aeed7ce0d95, status: pending). Contact details redacted; canonical application lives on filpgf.io. 1.1 Project Name Sentinel : Real-Time Behavioral Anomaly Detection for AI Agents 1.2 Project Github https://github.com/Harshit-Mishra2212/sentinel-agent-monitor 1.3 Project Website https://sunstone-project-page.vercel.app/roadmap.html 1.4 Team Lead/Point of Contact Harshit Mishra, Discord : harshitmishra0958, Telegram : @hm_22126 1.5 Category [ "RFP 3 - AI infrastructure products on Filecoin" ] 1.6 Open Source Status Fully Open Source 2.1 Project Summary AI agents are starting to do real things on Web3, signing transactions, calling contracts, moving funds around on their own. While doing this, they constantly produce background signals, how long things take, how much gas gets used, how the API responds. These signals can quietly reveal what an agent is actually doing, and nothing currently watches this in real time. Sentinel is an open source AI infrastructure tool for Filecoin based agent projects that watches agent behavior live, learns what looks normal, and flags it the moment something looks off. It also stores behavioral traces and anomaly flags on IPFS/Filecoin as a verifiable, tamper resistant audit log. The infrastructure Sentinel provides fills a gap that currently exists in the Filecoin AI agent stack, there is no open source tool teams can drop in to get real time behavioral monitoring plus verifiable on-chain logging of what their agent actually did. 2.2 Who does this work support? [ "Application Builders", "Other" ] 2.3 Total Funding Requested (USD) $26000 2.4 Milestones & Budget [ { "title": "Behavioral dataset generator and baseline benchmark", "description": "This is a 6 month plan starting September 2026. Milestone 1 is due by end of November 2026 (Month 3) and Milestone 2 is due by end of February 2027 (Month 6).\n\nProject start date: September 1, 2026. Total grant window: 6 months (September 1, 2026 to February 28, 2027). Total estimated hours: approximately 200 hours across the grant window. September and October may have some reduced availability due to campus internship season (aptitude tests, interviews), estimated at around 6 to 8 hours per week during those months. From November onwards availability increases to around 10 hours per week, which is why the more complex Filecoin integration work in Milestone 2 is scheduled for December through February\n\nBuild a dataset generator (Python CLI: generate_dataset.py --scenarios <list> --output <path>) that simulates AI agent behavioral telemetry, timing, gas usage, API call patterns, signing behavior, across a documented set of anomaly scenarios. Use this to benchmark the current detection model and publish baseline results. Budget covers development time (~$8,000, approximately 90 hours at roughly $100/hour), compute for training and benchmarking (~$2,000), and contingency (~$2,000, used only if additional training runs are needed to stabilize benchmark results, not discretionary). Minimum viable scope if time slips: dataset generator with at least 2 anomaly scenarios and partial benchmark results published.", "dueDate": "2026-11-30", "fundingRequested": "$12000", "completionCriteria": "generate_dataset.py published in the repo with a documented schema for each scenario type. Baseline benchmark report published in the repo showing the current model's accuracy on each scenario.\nAs part of Milestone 1, the GossipSub dispatcher-worker agent example already contributed to seetadev/py-libp2p will be instrumented to emit real telemetry, providing at least one real agent trace source included in the benchmark alongside the simulated dataset." }, { "title": "IPFS/Filecoin audit logging proof of concept and integration guide", "description": "Build a proof of concept that stores anomaly logs (JSON schema: timestamp, agent ID, anomaly score, behavioral snapshot) on IPFS/Filecoin, with a reproducible demo script for store, retrieve, and verify. Publish an integration guide showing how Filecoin based agent teams can adopt this pattern. Budget covers development time (~$10,000, approximately 110 hours at roughly $125/hour), Filecoin storage deal costs for testing (~$2,000, estimated at current deal pricing), and contingency (~$2,000, used only for additional deal testing if initial storage attempts fail or need retries). Minimum viable scope if time slips: working store/retrieve flow without full verification step, plus integration guide …[truncated] 3.1 Impact pathway This project directly targets the 2026 Network Objective: Drive Paid Onchain Deals. Sentinel provides AI infrastructure for Filecoin based agent projects, a behavioral dataset generator, anomaly detection benchmark, and IPFS/Filecoin audit logging proof of concept, filling a gap in the current Filecoin agent stack where no open source tool exists for real time behavioral monitoring with verifiable on-chain logging. The outputs to outcomes chain is: generate_dataset.py and demo.py ship as open source tools, Filecoin based agent teams adopt the integration guide to run Sentinel against their own agents, each agent running Sentinel stores behavioral logs as paid Filecoin mainnet deals, directly contributing to onchain deal volume. Anomaly logs are batched into bundles of 10 logs per deal, each bundle creating one paid storage deal. The 50 deals targeted during testing are minimal on-chain validation. The GossipSub dispatcher-worker example in seetadev/py-libp2p will be instrumented during Milestone 1 as the first external non-demo agent source. Target adopters include Filecoin based agent projects building on frameworks like Coinbase AgentKit integrated with py-libp2p, and teams building autonomous agents for onchain tasks who need verifiable behavioral records. 3.2 Verification metrics **1. Metric:** Storage deals generated from anomaly logs **Data source:** Filecoin network, storage deal records from the demo **How it's measured:** counting storage deals created during the proof of concept's logging activity **Target by end of grant:** demo running continuously enough to generate a documented number of real storage deals (e.g., 50+) during testing **2. Metric:** Detection accuracy on benchmark dataset **Data source:** project's own open source benchmark dataset **How it's measured:** comparing model output against known injected anomalies in the documented scenarios **Target by end of grant:** baseline results published in the repo **3. Metric:** Integration guide engagement **Data source:** GitHub repo (views, issues, discussions) **How it's measured:** tracking whether the published integration guide receives engagement (questions, forks, discussions) from other teams **Target by end of grant:** integration guide published, with at least 2 teams engaging with it (commenting, asking questions, or expressing interest) 3.3 References **Manu Sheel Gupta**, Developer, Learner, Contributor & Maintainer @ libp2p, multiformats, IPLD and SocialCalc/EtherCalc, working on LeakDetectAI and open source Web3 security tooling. 4.1 Monthly Operating Burn [ "< $10K (basic solo operation or part-time team)" ] 4.2 What % of total team monthly burn depends on this grant? 100%. There is currently no existing budget for this project. Time committed so far has been personal, unfunded time alongside coursework. 4.3 If this grant is not awarded, what happens? The proof of concept would remain at its current early stage. Without dedicated time, progressing it into a properly benchmarked, Filecoin integrated, documented tool would be very difficult to prioritize alongside coursework. 4.4 Core Team **Harshit Mishra**, 2nd year Computer Science student at NIT Kurukshetra, sole developer on this project. Background includes firmware development, full stack web development, hardware integration, and competitive programming, with prior open source contribution to py-libp2p (a Filecoin/IPFS related networking library). Currently building ML and Web3 specific skills through this project. Expected time commitment across the grant window September 1 2026 to February 28 2027: September 25 hours (internship season), October 25 hours (internship season), November 35 hours, December 35 hours, January 40 hours, February 40 hours, total approximately 200 hours. Manu Sheel Gupta is providing ongoing guidance on project direction and Web3 security context. Budget justification: Total hours across grant: approximately 200 hours. Effective blended hourly rate: approximately $90/hour across both milestones. Milestone 1: 90 hours at $90/hour = $8,100 rounded to $8,000. Milestone 2: 110 hours at $90/hour = $9,900 rounded to $10,000. Total dev time: 200 hours at $90/hour = $18,000. Contingency of $2,000 per milestone is triggered only by specific events: Milestone 1 contingency for additional training runs exceeding planned compute budget, Milestone 2 contingency for additional deal retries if initial storage attempts fail. Remaining $8,000 covers compute ($2,000), Filecoin deal testing ($2,000), and contingency across both milestones ($4,000 total). Filecoin deal parameters: anomaly logs will be batched into bundles of approximately 10 logs per deal, JSON format, estimated 50KB per bundle. Target: 25 deals on Filecoin mainnet at estimated current deal pricing of roughly $20-40 per deal, total estimated cost $1,000-$2,000 covered by the deal testing budget. All deals will be paid mainnet deals to directly contribute to Drive Paid Onchain Deals objective. Budget to impact summary: $26,000 total across 200 hours produces two milestones. Milestone 1 delivers a dataset generator and benchmark report. Milestone 2 delivers a store/retrieve/verify demo plus integration guide, generating at least 25 paid mainnet storage deals (approximately 250 anomaly log entries) during testing, funded partly by the $2,000 Filecoin deal testing budget. $26,000 is the right ask for this grant because it covers a focused, solo, student-led effort to build and validate the core infrastructure primitive, a behavioral dataset generator, anomaly detection benchmark, and Filecoin logging proof of concept, before scaling to a larger team or broader deployment. A smaller ask would not cover the Filecoin deal testing costs and compute required to produce credible benchmark results. A larger ask is not justified at this stage without external pilot commitments already in place. Prior open source contribution to py-libp2p : https://github.com/seetadev/py-libp2p/pull/45 , Sugar Labs M …[truncated] 4.5 Has your team received a ProPGF grant or funding from PLFIF before? [ "No" ] 5.1 Key risks & dependencies As a solo student contributor working around 10 hours per week, availability is reduced during exam periods and internship season which could affect timelines. The project depends on continued learning of ML concepts as it progresses. Technical risk includes the model's performance on more realistic data potentially requiring further architecture adjustments beyond what's currently scoped. The Filecoin/IPFS storage integration is a new component for this project and may take longer than estimated to get working reliably. Any feedback you have on the application process? The form is thorough and the milestone structure is clear, which is helpful for thinking through the project realistically. Anything else you want to share that we didn't ask? This project grew out of a broader interest in real-time anomaly detection, inspired by my work on a separate IoT health monitoring project that uses a similar approach for detecting medical emergencies before they happen. Applying that same core idea to AI agent behavior on Web3, combined with using Filecoin/IPFS for verifiable audit trails, felt like a natural and underexplored direction. Objective 1 Direct Objective 2 N/A Objective 3 N/A