Mirrored from filpgf.io — ProPGF Batch 3 (Karma program 1479, application 6a1893345513a8d9decf00d6, status: pending). Contact details redacted; canonical application lives on filpgf.io. 1.1 Project Name InferGrid: Verifiable AI Work Receipts And Agent Memory On Filecoin 1.2 Project Github https://github.com/kernelogic/InferGrid 1.3 Project Website https://infergrid.filweb3.com 1.4 Team Lead/Point of Contact Fei Yan, Founder / Lead Developer. Preferred contact: email at [redacted] or Filecoin Slack. 1.5 Category [ "Core Infrastructure", "RFPs (Coming Soon)" ] 1.6 Open Source Status Partial 2.1 Project Summary InferGrid will build a managed verifiable AI work service for embeddings, RAG, and agent workflows on Filecoin. Customers submit AI jobs through a familiar API; large model artifacts, datasets, RAG indexes, agent memory, and task receipts are represented through CID-based manifests and stored or referenced through Filecoin-compatible paths. Provider nodes retrieve and verify these artifacts before execution, and completed work produces tamper-evident receipts and audit bundles. This work supports AI teams that need provenance, Filecoin storage providers, retrieval operators, application builders, and ecosystem teams by turning Filecoin storage and retrieval primitives into a practical customer-facing AI infrastructure product. 2.2 Who does this work support? [ "Storage Providers", "Application Builders", "Application Users", "Network Infrastructure", "Onramps", "Pods" ] 2.3 Total Funding Requested (USD) 200000 2.4 Milestones & Budget [ { "title": "Customer-Facing Artifact And Receipt API", "description": "Define and implement the first customer-facing Filecoin-backed artifact and receipt API for AI workloads. Deliverables include a CID-based manifest format for model weights, adapters, quantized artifacts, datasets, RAG indexes, and task receipts; a buyer-facing API path for one focused embeddings, RAG, or agent-workflow use case; a reference uploader or integration path for storing public test artifacts through Filecoin Onchain Cloud-compatible storage; an integrity verification flow that provider nodes can run before accepting work; and public documentation for artifact publishers and AI builders.", "dueDate": "2026-07-31", "fundingRequested": "40000", "completionCriteria": "A customer-style API call can submit a test workload and receive a receipt. A public demo artifact can be registered, retrieved, hash-checked, and used by a provider-node workflow. Documentation explains how a third party can publish and verify an artifact." }, { "title": "Retrieval-To-Compute Provider Flow", "description": "Let provider nodes retrieve Filecoin-backed artifacts and prepare them for customer-visible AI work. Deliverables include a provider-node flow for resolving an artifact manifest, retrieving required files, verifying integrity, and caching locally; a basic cache policy for large model/data artifacts; receipt fields that prove which artifact version was used for a task; and an SP participation guide covering storage, retrieval, cache, preprocessing, and optional compute roles.", "dueDate": "2026-08-31", "fundingRequested": "50000", "completionCriteria": "A provider node can fetch a Filecoin-backed model or dataset artifact and use it in a controlled embeddings, RAG, or agent-workflow test. The resulting task receipt references the artifact identity and verification result." }, { "title": "Audit Evidence And Fee-Surface Instrumentation", "description": "Make Filecoin-backed audit evidence and future FIL/Filecoin-compatible value-accrual surfaces visible. Deliverables include an audit bundle format that links customer job, artifact identity, provider verification, receipt, and Filecoin-backed storage/retrieval evidence; prototype instrumentation for artifact storage, retrieval events, receipt archival, retention period, and potential fee surfaces; a design note for future FIL/Filecoin-compatible settlement, burn, or lock mechanisms tied to measurable network activity; and risk notes for custody, failed work, duplicate receipts, provider disputes, manual review, and avoiding premature mainnet claims.", "dueDate": "2026-09-30", "fundingRequested": "55000", "completionCriteria": "A controlled demo produces a customer-inspectable audit bundle for completed work. Failed or unverifiable work does not produce a valid completion receipt. …[truncated] 3.1 Impact pathway InferGrid creates a concrete AI workload path for Filecoin storage and retrieval primitives. The output is a customer-facing artifact, receipt, and audit layer: AI teams submit embeddings, RAG, or agent work; provider nodes retrieve and verify Filecoin-backed artifacts; and completed work produces customer-inspectable audit evidence. This gives storage providers and retrieval operators more AI-adjacent demand, including storage, cache, preprocessing, verification, receipt archival, and optional compute roles. As these workflows mature, they can increase paid storage and retrieval activity, improve provider utilization, and make Filecoin more useful to AI application builders. 3.2 Verification metrics 1. Public AI artifact manifests created Data source: public repo, docs, and demo logs. Measured by counting valid manifests for model, dataset, RAG index, or receipt artifacts. Target: at least 3 sample artifact manifests by the end of the grant. 2. Provider retrieval and verification demo Data source: demo logs and receipt output. Measured by showing a provider node resolving a manifest, retrieving the artifact, verifying integrity, and recording the result. Target: 1 end-to-end provider demo. 3. Task receipts with artifact provenance Data source: receipt JSON or public demo output. Measured by confirming receipts include artifact CID or manifest identity plus verification status. Target: at least 5 sample receipts. 4. Customer audit bundles created Data source: demo output and public docs. Measured by showing completed work produces a receipt bundle with artifact identity, verification status, and Filecoin-backed evidence. Target: at least 5 sample audit bundles. 5. Filecoin-backed activity instrumentation Data source: demo logs and public metrics. Measured by counting stored artifacts, retrieval events, receipt archival, and retention surfaces. Target: 1 public metrics view or report. 6. SP/operator documentation Data source: public documentation. Measured by publishing a guide for storage, retrieval/cache, CPU/storage-heavy, and GPU-capable participation modes. Target: 1 complete operator guide. 3.3 References fil+ notary singularity v1 founder 4.1 Monthly Operating Burn [ "$10-$100K (small team)" ] 4.2 What % of total team monthly burn depends on this grant? 100% 4.3 If this grant is not awarded, what happens? Without this grant, InferGrid will likely continue the general AI compute marketplace work, but the Filecoin-first artifact, retrieval, receipt, and audit-bundle path will be delayed or reduced to planning only. The grant accelerates the customer-facing product surface that directly benefits Filecoin storage providers, retrieval operators, and AI builders. 4.4 Core Team Fei Yan leads product direction, architecture, and implementation for InferGrid. The initial delivery plan is intentionally small and milestone-based: artifact manifest design, provider-node retrieval and verification, settlement adapter design/prototype, SP/operator documentation, and a public demo. Additional ecosystem or operator support can be added around the SP pilot as the grant scope is confirmed. 4.5 Has your team received a ProPGF grant or funding from PLFIF before? [ "No" ] 5.1 Key risks & dependencies Key risks are integration complexity, evolving Filecoin Onchain Cloud/payment APIs, audit-evidence correctness, settlement safety, and scope control. The project reduces risk by starting with public test artifacts, controlled testnet/demo evidence, explicit receipt formats, and milestone-level verification before any production mainnet exposure. Another risk is overclaiming storage-provider compute availability, so the SP plan separates storage, retrieval/cache, CPU/storage-heavy, and optional GPU roles. Private marketplace algorithms and enterprise features are intentionally out of scope so the grant remains focused on a usable Filecoin-backed AI infrastructure product. Any feedback you have on the application process? No feedback at this time. Anything else you want to share that we didn't ask? This proposal is designed as Filecoin-first work, not a generic AI marketplace grant. The public deliverables focus on a customer-facing AI infrastructure product and reusable Filecoin ecosystem value: buyer-facing API path, artifact manifests, retrieval-to-compute verification, receipt provenance, audit bundles, Filecoin activity instrumentation, SP participation docs, and a testnet demo. The commercial marketplace strategy remains outside the grant scope. Contributing to Core Infrastructure? InferGrid builds customer-facing AI infrastructure on Filecoin: artifact manifests, retrieval-to-compute verification, task receipts, and audit bundles for embeddings, RAG, and agent workflows. AI builders depend on it for provenance and verifiable delivery; Filecoin storage providers and retrieval operators benefit from new demand for artifact storage, retrieval, cache, and receipt archival. Objective 1 Indirect Objective 2 Indirect Objective 3 Indirect Open Source Context The grant-funded outputs will be public where they create Filecoin ecosystem value: buyer-facing API shape, artifact manifest format, sample manifests, retrieval and verification flow, task receipt fields, customer audit-bundle format, Filecoin activity instrumentation, SP/operator documentation, and a public testnet demo. Some product-specific marketplace components may remain private, including provider routing heuristics, abuse controls, provider scoring, enterprise pipeline, and commercial pricing experiments. This boundary keeps the Filecoin integration inspectable and reusable while avoiding disclosure of unrelated proprietary marketplace strategy.