AI Consultant

We are running a two-milestone proof of concept and are looking for a hands-on consultant who can stand up self-hosted LLM infrastructure and build a scoring service on top of it. This is scoped, deliverable-driven work with a clear end point — not an open-ended role. The right person has done both of these things before and can move quickly with minimal supervision. The goal of the POC is to identify the resource with the right skill set. Cloud accounts, GPU budget, and sample data will be provided by us. Milestone 1 — vLLM deployment on cloud infrastructure Stand up a working, self-hosted inference environment. Scope Provision and configure vLLM on cloud-hosted GPU servers ([VastAI]) Deploy and serve multiple LLM models of varying sizes and architectures; advise on GPU sizing, quantization, and memory/context configuration for each Expose an OpenAI-compatible API endpoint and demonstrate querying the models from a remote client machine outside the host network, with appropriate auth and network controls Support model switching / multi-model serving and document the tradeoffs of the chosen approach Deliverables Running vLLM environment with the agreed models deployed Basic documentation Milestone 2 — Browser fingerprint comparison and scoring service Build a service that compares a browser fingerprint against a history of prior fingerprints and returns a score. Scope Accept a JSON payload containing a current browser fingerprint plus an array of 0–20 historical fingerprints Perform element-by-element comparison across fingerprint attributes, handling variable array length (including an empty history), missing or partial attributes, and differing schema versions gracefully Produce a single score reflecting how consistent the current fingerprint is with the history, plus a per-attribute breakdown showing what drove the result Make attribute weighting configurable rather than hard-coded, so scoring logic can be tuned without a code change Build a wrapper API service (FastAPI) exposing the scoring endpoint, with request validation, error handling, and OpenAPI documentation Where useful, incorporate an LLM or agent-based component — running against the Milestone 1 infrastructure — to reason over ambiguous or partial attribute differences Deliverables FastAPI service with documented endpoints and a published OpenAPI spec Scoring implementation with a configurable weighting scheme and a written explanation of the methodology Test suite covering edge cases: empty history, maximum array size, malformed input, contradictory attributes Sample requests and responses, and a short guide on interpreting and tuning scores Required skills and experience Deep, hands-on LLM deployment experience — serving open-weight models in production or production-like environments, including GPU sizing, quantization, batching, and context/memory configuration Use AI tools like Claude or ChatGPT (paid veersion) to generate code Direct hands-on experience with vLLM — not just familiarity; you have deployed, configured, and tuned it Strong Python and FastAPI — designing and shipping clean, documented API services Experience developing agents — agentic workflows, tool use, and orchestration against LLM backends Cloud and infrastructure fundamentals — GPU instance provisioning, containerization, CUDA/driver setup, networking, and securing remote access to inference endpoints Independence — able to scope, execute, and hand off documented work with light oversight and clear written communication Nice to have Background in browser fingerprinting, device intelligence, bot detection, or fraud/risk scoring Experience with similarity scoring, fuzzy matching, or weighted heuristic models Familiarity with model gateways and serving layers (LiteLLM, Ray Serve, Kubernetes), plus inference observability Track record of taking a POC through to a production hand-off Acceptance criteria Each milestone is considered complete when the deliverables above are handed over, the environment or service runs from a clean setup following the provided documentation, and a walkthrough session has been completed with our team.