Bioinformatics Data Scientist / Engineer — Microbiome Knowledge Graph

⚠️ Read before applying: I will only set up calls with shortlisted developers who provide a fixed price for each milestone first. Proposals without fixed quotes won't be shortlisted. Overview We're building the evidence engine behind a consumer + practitioner gut microbiome testing product (shotgun NGS). We have a structured knowledge-graph database mapping microbial markers to health conditions, interventions, and complementary tests, and we need it populated with high-quality, qualifying scientific evidence — then automated. The engagement is two milestones, each its own fixed price: Milestone 1 — MVP-launch backfill (~5,000 qualifying pieces). The dataset we need to launch. Also proves out the approach and quality. Milestone 2 — Full automation pipeline. Built after launch (so it never blocks it): the ongoing production system for continuous ingestion and scaling to many more niches. Milestone 1 is the priority — we launch on it. Milestone 2 is a go/no-go you scope once M1 delivers. The tooling you build for M1 carries directly into M2. The full database schema, node taxonomy, and scoring framework are proprietary and shared under NDA. Milestone 1 — the backfill (what "populating" means) Our database is a knowledge graph — evidence doesn't just attach to fixed entries, it also generates new ones. This milestone spans four layers, each needing its own qualifying evidence: Backfill existing nodes — attach graded marker↔condition evidence to nodes already in the graph. Discover & add new nodes — searching the literature surfaces markers we don't yet have; those become new nodes and get populated. Intervention evidence — each food/supplement/drug that shifts a marker needs its own interventional (RCT-grade) support. Complementary-test evidence — each external lab test we pair with a finding needs a convergent-validity (correlation) study. Target: ~5,000 unique qualifying pieces, broken down: ~3,000 (60%) — marker↔condition node evidence ~1,000 (20%) — intervention evidence ~650 (13%) — complementary-test evidence ~350 (7%) — discovery headroom for new nodes/layers We measure and pay against qualifying evidence, not raw articles screened. Qualifying = deduplicated, on-topic, direction verified, human-relevance prioritized (animal/mechanistic capped as emerging), study design tiered, conflicts of interest flagged. Delivery = the qualifying set plus an honest gap register (what didn't qualify and why). Can be delivered in two ~2,500 tranches to de-risk funding on both sides. Milestone 2 — the full automation pipeline Productionize the approach into an efficient, monitored, scalable pipeline (full spec shared under NDA): Source search across PubMed, Europe PMC, OpenAlex, Semantic Scholar, bioRxiv/medRxiv Embedding-based relevance pre-filter (spend the expensive step only on on-topic papers) LLM extraction to a strict structured schema — run on a Claude subscription "harness" (MCP/agent on a paid seat, headless), not the metered API, to control cost; metered API as overflow fallback Entity normalization to canonical IDs (NCBI Taxonomy / GTDB), reclassification handling Deduplication + deterministic scoring → evidence tier Full-text capture ~90% via: PMC OA, Europe PMC OA, Unpaywall, preprints, publisher text-mining (TDM) APIs, and licensed/institutional access for the tail. Playwright to aggreagte data on Sci-Hub to close the gap on the 90%. Human approval queue — one review queue for edge cases, ambiguous entity merges, direction conflicts, conflict-of-interest calls, and new-ontology approvals. Nothing goes live without human sign-off. Health-monitoring dashboard — live health checks across the pipeline: per-stage throughput/errors, source status and full-text capture rate vs the ~90% target, harness quota/cost, data-quality metrics, and coverage/milestone progress, with alerting. Must-haves Biomedical / epidemiology literacy: can judge study design, human vs animal relevance, direction, confounders (e.g. medication), and conflicts of interest Experience building literature-ingestion / NLP extraction pipelines Comfort with the research APIs above and structured data output Entity resolution / normalization experience Clean, organized, scalable, well-documented code (see terms) Nice-to-haves Microbiome / gut-health domain knowledge Knowledge-graph / ontology experience Human-in-the-loop review workflows and pipeline observability / monitoring dashboards Playwright / headless browser automation (legitimate rendering) Prior work with Claude / MCP agents or similar LLM orchestration Terms Fixed-price, per milestone. Quote a separate fixed price for Milestone 1 (the ~5,000-piece backfill) and Milestone 2 (the full automation pipeline). You may deliver Milestone 1 with your own tooling — and it carries forward. Any pipeline/code you build to produce the M1 backfill is delivered to us as work-for-hire and reused in Milestone 2, so you won't rebuild it and we won't pay for it twice. Price M1 as the evidence-production effort on top of that reusable tooling, not a throwaway one-off build. Milestones are escrow-funded and released on delivery against the agreed criteria (M1: qualifying set + gap register; M2: working pipeline per spec). M1 may be funded/delivered in two tranches. All work is reviewed by an independent 3rd party before each milestone is closed out. Your code and outputs must be organized, clean, scalable, and documented enough to pass that review and be handed off/extended. A brief kickoff alignment call with our project lead sets scope and review standards before Milestone 1 begins. Async / written-first collaboration is welcome after that. How to apply Quote your fixed price for each milestone (M1 backfill, M2 automation) plus cost assumptions. I only shortlist and schedule calls with developers who provide fixed quotes first — no exceptions. When quoting, assume the tooling you build for M1 is reused (not rebuilt) in M2 — price accordingly. Share a similar pipeline you've built (link or description). Explain how you'd handle the recall-vs-precision tradeoff and reach ~90% legal full-text capture. Give a rough plan for one qualifying bundle so we can measure real cost/coverage. Application test (please answer) Once you have access to the worksheet under NDA, how would you go about populating it comprehensively? Walk us through your understanding end-to-end — where you'd start, how you'd know what evidence to search for, how you'd find and fill gaps (including whole categories that are light or empty), and how you'd keep extraction accurate and the data clean at scale. Screening questions Describe an evidence/literature-ingestion pipeline you've built — its accuracy and throughput. How do you keep LLM extraction structured and verifiable at scale? Have you run Claude/LLMs via a subscription harness vs the metered API? How would you control cost here? How would you design the human-in-the-loop review queue and the pipeline health-monitoring/alerting? NDA The full database schema, node taxonomy, category structure, and scoring framework are shared after a signed NDA. Happy to provide a redacted sample and answer scoping questions beforehand.