Rishabh Patil: models, data and research

AI engineer and researcher in London. Co-founder and CEO of Valuren, which gives fashion and luxury brands digital product passports. I work on verification for AI: checking claims against evidence, evaluating LLM systems, and planning safely in learned latent spaces.

On the Hub

Downloads

Loading live numbers

2,788

downloads, all time

+308 in the last 30 days

models

dataset

Space

Most downloaded repos

Hover or tab through the bars for 30-day numbers

Table view of every repo

Repo	Type	All time	Last 30 days

VeriSci: scientific claim verification

Paste a scientific claim and get SUPPORTS, REFUTES or NOT ENOUGH INFO, with the sentences that justify the label.

VeriSci pipeline: claim, retrieve with BM25 and a fine-tuned e5 retriever, select evidence sentences, verify with a calibrated DeBERTa model, return a label with evidence. Retriever Recall@5 0.867. Verifier accuracy 0.905, macro-F1 0.874, calibration error 0.024. End to end on SciFact validation: 0.691 accuracy, 0.653 macro-F1.

Live demo (may take a minute to wake if idle)
Release verifier and retriever
Seeded ablations of evidence selection, each with its own model card

SchemaSage-SQL: safe text-to-SQL

QLoRA adapters on Qwen3-4B that write read-only SQL grounded in the schema you give them, with a safety layer that refuses destructive requests.

SchemaSage-SQL release baseline on 64 cleaned held-out examples: 100% SQL parse validity, 98.3% schema adherence, 0% unsafe queries, 100% correct refusals. QLoRA on Qwen3-4B-Instruct-2507, trained on 111,444 cleaned examples, 10,862 of them refusals.

Release baseline; the larger candidate scored higher on exact match and execution accuracy but missed one blocked refusal, so it was not shipped
Dataset and code

Research code

Each library pairs installable code with a first-author preprint.

SASP results: test cross-entropy over 120 epochs against the baseline, and Hessian trace at convergence, about 171 for the baseline against about 100 with SASP.

toploss

Five optimiser-free PyTorch regularisers that rebuild SAM, cautious weight decay and AdEMAMix as loss penalties. SASP cut the Hessian trace 41% at +15% wall-clock, against +107% for SAM (three seeds, CPU).

Preprint

CLAP planning in a two-dimensional latent field: the trajectory reaches the safe target region and dwells there, avoiding a reward trap and an unsafe zone.

clap-family

Conservative Lapse-Action Planning: reach the best safe region of a latent space, then stay there. Seven planner variants, 72 tests.

Preprint

micro-world-model · preprint
Hierarchical JEPA world model that plans in a 256-dimensional latent space with no reconstruction, no reward and no EMA target.
ariadne-eval
Trajectory-level tracing and scoring for multi-step, tool-using agents. 285 tests, mypy-strict.
evalforge
LLM evaluation with bootstrapped confidence intervals, calibrated LLM-as-judge and a CI regression gate. 538 tests.

Background

MSc Artificial Intelligence, University of St Andrews (dissertation on dependent types for machine learning, supervised by Dr Edwin Brady). BSc Artificial Intelligence, VU Amsterdam. Before Valuren I co-founded RelAIable, an AI consultancy in Amsterdam, as Chief AI Officer.