Item 65 · edge

Input

Theo Davenport
Head of AI at Clauselink

I came to AI through a fairly traditional NLP research path. I did my PhD at Cascadia State University on information extraction from semi-structured documents, then joined Helix Labs in Westchester for three years, where I worked on entity resolution pipelines for financial regulatory filings. The work was rigorous and I learned a lot about what production NLP actually requires — namely, that precision matters more than recall when the downstream consumer is a human who has to act on the output.

From IBM I moved to a Series B startup called Contegra, which built compliance automation tooling for insurance carriers. I led the NLP team there, growing it from two engineers to eight over two years. We built a clause-level extraction system on top of BERT-era models, and I made the call early to use ensemble methods rather than fine-tuning a single large model, mostly because our labeled data was thin and the ensembles were more auditable — something insurance clients cared about deeply. We got acquired by a larger RegTech firm in 2021, and I stayed through the integration for about eight months before deciding the pace wasn't right for me.

After Contegra I spent eighteen months at Verdigris AI as a member of technical staff focused on document understanding applications, which is where I got deep exposure to the current generation of large language models and their failure modes in structured extraction tasks. I worked closely with the evals team to build benchmarks for multi-page document comprehension, and that experience shaped how I think about measuring AI quality in domains where ground truth is genuinely ambiguous — something that comes up constantly in legal text.

I joined Clauselink about two years ago because the contract review problem is one of the hardest document AI problems I know of. Legal language is adversarially ambiguous by design. My team currently has four ML engineers, two research engineers, and a data annotation lead. We run a hybrid architecture: a retrieval-augmented pipeline handles clause identification and cross-reference resolution, while a fine-tuned LLM layer handles risk classification and deviation flagging against customer-specific playbooks. The hardest ongoing challenge is maintaining accuracy as customer playbooks diverge — we've been investing heavily in per-customer adapter layers rather than full model retraining, which has meaningfully reduced our update latency.

My current focus is on two things: getting our confidence calibration to a point where the product can surface reliable uncertainty signals to reviewers, and expanding coverage from MSAs and NDAs into more bespoke agreement types like JVs and IP licensing. Both require better training data than we have, so a non-trivial part of my time goes into annotation infrastructure and quality control.

Outside of Clauselink, I advise two early-stage startups in the document intelligence space, I contribute intermittently to the open-source RetrieveStack project, and I speak at legal tech conferences a few times a year, mostly on the gap between benchmark performance and real-world production behavior in legal AI systems.

Clauselink is a B2B SaaS company focused on AI-powered contract review for enterprise legal and procurement teams. The platform enables legal professionals to extract, classify, and flag risk across commercial agreements at scale, with support for customer-specific playbooks that reflect individual negotiation standards. Clauselink's AI layer combines retrieval-augmented extraction with fine-tuned classification models, providing reviewers with structured risk summaries and deviation alerts rather than raw model output. The company serves mid-market and enterprise customers in sectors including technology, financial services, and professional services. Clauselink has raised a Series C round and is expanding its coverage beyond standard agreement types into more complex deal structures.

Gold

Expected action

auto_add

Fit score

0.92

Industry

B2B SaaS

Segment

Legaltech / AI contract review

Seniority

Director

Company size

—

Language

en

Notes

Very long bio but rich with structural signals. Headcount not explicitly stated in profile or company; Series C scale supports in-range inference but a labeller could ask for verification.

Predictions