run real instruments.

Grounded literature assistant over a research PI's corpus. Resolves the lab on OpenAlex, ingests open-access full text, hybrid dense+BM25 retrieval, answers with citations.

Predict ΔΔG of point mutations; deep-mutational scan a position.

ML structural / disorder / feature analysis from a sequence or UniProt accession.

13 ADMET models over a SMILES library; ranked drug-likeness report.

Mine molecular-dynamics trajectories for conformational structure (demo trajectory until GPU compute lands).

ML over patch-clamp electrophysiology recordings; cell-type signatures.

Triage cryo-EM micrographs for quality (synthetic session until GPU compute lands).

Personalized recent-paper feed. Queries the live OpenAlex index for your topics, ranks by relevance + recency + citation velocity, and explains why each matters to you.

Funder finder + specific-aims drafter, grounded in real awarded NSF grants (research-atlas corpus). Shows who funds your area and drafts aims anchored to actual awards.

Which method answers your question? Mines the recurring methods in the live OpenAlex literature and points you to the Bucket tool that runs it.

Cross-paper consistency check. State a claim; ReviewGuard sorts the OpenAlex literature into supporting vs contradicting, quoting the deciding sentence.

Evidence, not hype. State a quantum-biology claim; QuantumBioRAG scores how strongly the live OpenAlex literature supports it — weighting each paper by overlap, citations, and recency — with a consensus score and the deciding sentences.

RNA secondary-structure prediction via ViennaRNA: MFE dot-bracket structure, free energy, partition-function base-pair probabilities, and a readable helix/loop summary. Fully real thermodynamics.

CRISPR SpCas9 guide design: PAM scan on both strands, transparent on-target efficiency scoring, and a local seed-region off-target risk flag. Ranked, defensible guide table.

RNA → ML embedding. Real RNA-FM language-model representation when its weights are installed; otherwise an honest, reproducible k-mer + structural-feature embedding (mode reported).

Fit passive-membrane (RC) parameters — R, C, τ, V₀ — to a current-clamp trace via scipy least-squares, with fit quality (R²/RMSE). A demo trace with known params verifies recovery.

Detect spikes in a voltage trace (MAD-robust threshold + refractory + alignment) and extract real waveform features — amplitude, width, half-width, firing rate, ISI stats. Demo train has a known spike count.

Methods prose → a structured, runnable protocol. Deterministic rule extraction over a methods knowledge base: ordered steps with timings/temps/volumes, a reagent table, and safety flags. No network, no GPU.

Map a toxin/peptide (name or sequence) to its likely ion-channel targets. Fuses a curated venom-peptide pharmacology KB with live OpenAlex co-occurrence; sequences classified by cysteine framework. Ranked targets, honest confidence, cited exemplars.

Build a paper's local citation neighborhood from the live OpenAlex graph (DOI / OpenAlex ID / title). Surfaces the key related works and ranks them by degree centrality — the most-connected neighbors first.

Calcium-imaging ΔF/F + event detection. Real rolling-percentile F0 baseline, ΔF/F, MAD-robust transient detection with single-exponential decay-τ fits, and firing-rate stats. Demo trace has a known event count.

Cell / nuclei segmentation. Cellpose on CPU when installed; otherwise a real classical pipeline — Otsu threshold + distance-transform seeded watershed — with per-object area, centroid, and bounding box. Demo image has a known cell count.

AFM force-curve analysis. Real contact-point detection + Hertz (sphere) / Sneddon (cone) least-squares fit for Young's modulus in SI units, plus adhesion from the retract minimum. Demo curve has a known modulus.

Traction-force / PIV displacement field. Real block-matching (normalized cross-correlation) between a relaxed and a deformed bead image → per-window vectors + a strain-energy proxy. Classical, honestly labelled. Demo recovers a known shift.

Mine a paper's figures + stats. Real text-layer extraction of figure/table captions, reported statistics (p-values, n=, CIs, R²/r, mean±SD, fold-change), and unit-bearing measurements, linked per-figure. PDF or pasted text; pixel-level plot digitization is a documented GPU/vision extension.

DNA accessibility / regulatory-potential from sequence. Real interpretable feature model — GC content, Gardiner-Garden CpG islands, core-promoter motif scan (TATA / GC-box / CAAT / Initiator) — into a 0–1 accessibility score. A deep DNA-LM (Enformer/Evo) is the documented GPU path.

Amyloid / aggregation propensity from a protein sequence. Real windowed model — Chou-Fasman β-sheet propensity + Kyte-Doolittle hydrophobicity − net charge — flagging contiguous aggregation hot-spots. Interpretable and deterministic.

Single-channel idealization. Real half-amplitude threshold idealization of a single-channel current record into open/closed states, with open probability, dwell-time histograms, and ML single-exponential dwell constants. Demo recovers a known open probability.

Score a dataset's metadata for FAIR compliance. Real deterministic rubric over Wilkinson 2016's 15 sub-principles — persistent identifier, open license, machine-readable/standard formats, community vocabularies, provenance — into per-principle subscores, an overall 0–100 FAIR score, and a prioritized fix list. Funder-mandated (NIH/NSF/Horizon/Wellcome/Gates DMSP).

Check reported statistics for reproducibility. Real statcheck-style p-value recomputation (t/F/χ²/r + df → exact two-tailed scipy), the GRIM test for impossible means, and flags for missing multiple-comparison correction, CIs, and effect sizes. Paste a Results section; deterministic, never guesses.

Design a causal study (econ/social). Describe treatment, outcome, confounders, and the assumed causal graph; CausalDesigner builds the DAG, enumerates the backdoor paths, finds a valid minimal adjustment set via real do-calculus (Pearl's back-door criterion + networkx d-separation — it never conditions on colliders or mediators), and recommends an estimator (DiD/RDD/IV/matching/regression) with its identifying assumptions and threats to validity.

Featurize a composition for materials-property ML. Parse a chemical formula (subscripts, fractions, nested parentheses) and compute real Magpie-style descriptors (Ward 2016) — composition-weighted mean/range/average-deviation of atomic weight, Pauling electronegativity, atomic radius, melting point, valence-electron count — from a built-in element table. A ready-to-model feature vector, no GPU.

Statistical power & sample-size calculator (universal; esp. social/biomed). Solve for n, power, the minimum detectable effect, or alpha for the two-sample/one-sample t-test, one-way ANOVA (Cohen's f), two proportions, and Pearson correlation — real closed-form scipy noncentral-distribution power (the G*Power equations), recovering the textbook n=64 for a medium two-sample effect. A priori power planning is funder/IRB-expected.

Summarize a geospatial / time-series dataset (earth-climate). Descriptives + missing-data accounting, trend via OLS AND the distribution-free Mann-Kendall test + Theil-Sen slope (the standard climatological trend test), per-phase seasonal climatology + variance explained, lag-1 autocorrelation, and spatial extent (bbox/centroid/haversine span). Real numpy/scipy, reproducible, for non-specialists.

Score an ML experiment's reproducibility + emit a model card (cs-ml). Describe the dataset, splits, seed, hyperparameters, training, compute, eval, and what you shared; MLReproCard scores a real weighted rubric (NeurIPS/ICML checklists, Mitchell 2019 Model Cards, Gundersen's taxonomy) across data/code/training/evaluation/compute/sharing, flags exactly which repro elements are missing, assigns an R0–R3 level, and fills in a normalized model card. Deterministic, no LLM.

Pairwise sequence alignment (biomed-bio, the largest field). Exact dynamic programming — Needleman-Wunsch global and Smith-Waterman local — with the real BLOSUM62 substitution matrix for proteins (or an identity matrix for nucleotides) and a linear gap penalty. Returns the aligned strings, optimal score, match/mismatch/gap counts, and percent identity. The bedrock operation of every sequence pipeline, no heuristics, no GPU.

Balance a chemical equation + stoichiometry (chemistry). Balancing is solved as a null-space problem on the element matrix via exact rational Gaussian elimination, scaled to the smallest positive integers and re-verified element-by-element (H2 + O2 → H2O gives 2, 1, 2). Supply reactant amounts and it computes the limiting reagent, extent of reaction, and theoretical product yields. Real linear algebra, no lookup tables.

Dimensional analysis, unit conversion & equation consistency (universal; physics/engineering). Parses unit expressions into exact rational exponents over the 7 SI base dimensions, converts between compatible units (with affine °C/°F), and checks that an equation is dimensionally homogeneous — catching F = m·a (consistent) vs a wrong F = m·v (mass·length·time⁻¹ ≠ force). The cheapest correctness check in physical science, zero dependencies.

Survival / time-to-event analysis (biomed + econ-social). The exact Kaplan-Meier product-limit estimator with Greenwood standard errors and median survival, per group, plus the Mantel-Cox log-rank test (hypergeometric expected/variance, χ² on 1 df via scipy) between two groups. Handles right-censoring. The workhorse of trials, epidemiology, and event-history models — real, reproducible, no GPU.

Decompose & forecast a time series (econ-social, earth-climate, universal). Holt-Winters triple exponential smoothing (level/trend/seasonal, α/β/γ fit by minimizing in-sample SSE) with classical additive decomposition, and — the honest part — a holdout backtest reporting MAE/RMSE/MAPE against a naive last-value baseline. Dependency-light real numpy (statsmodels if installed); a forecast that doesn't beat naive adds nothing.