UK startups I'd want to work with
I put this together because the most useful thing an AI deployment engineer brings to a first conversation is a hypothesis about where the work is. These are early-stage UK companies building AI-native products; for each, my guess at the hard problem in their core AI workflow and the first eval I'd write. Every hypothesis is exactly that — I have no inside knowledge, and nothing here says or implies anything about anyone's customer or vendor relationships. Read it as my engineering judgement, and disagree with it in an interview.
Legal & property workflows
Wordsmith
AI platform for in-house legal teams — contract drafting, review and research. [source]
The hard part is likely answering from a live contract repository where amendments and side letters silently change effective terms — the exact failure I reproduce on page 1.
I'd write a stale-amendment eval over a synthetic repo of agreements plus amendments, and measure how often answers track the effective state rather than the original text.
Orbital
Property due-diligence and risk tech for real-estate lawyers. [source]
I'd expect false presence to be the dangerous mode: title documents are long, scanned and inconsistent, and a confident answer about a covenant that isn't there is worse than no answer.
I'd build an absent-provision eval on synthetic title-register-style documents and measure abstention calibration under retrieval noise.
Definely
Structured review tools for complex contracts, inside Word. [source]
The hard problem is likely defined-term and cross-reference resolution — "as defined in Schedule 3" chains where two near-identical definitions differ in one material word.
I'd probe near-duplicate defined terms: an eval where nested definitions collide, graded on whether the tool resolves to the controlling one.
Lawhive
Online consumer legal services delivered with licensed solicitors. [source]
I'd expect the AI-shaped risk to sit in intake and triage — routing a messy consumer description to the right legal matter type, with a hard boundary before anything resembling advice.
I'd instrument the intake flow and write a misrouting eval on synthetic matter descriptions, starting with the categories where a wrong route is costliest.
Regulated operations
Gradient Labs
AI agents for customer operations in financial services — disputes, KYC, support. [source]
The hard part is likely procedure adherence across long multi-turn conversations: single-message evals won't catch an agent that drifts off a mandated process at turn nine.
I'd build a multi-turn eval harness with synthetic personas (including vulnerable-customer scenarios) graded programmatically against the procedure the agent was meant to follow.
Sprout.ai
AI claims automation for insurers — coverage checking and fraud detection. [source]
Policy coverage checking is contract QA under a different name — I'd expect near-duplicate per-peril limits and endorsements that supersede base wording to be the recurring trap.
I'd re-shape my amendment eval around synthetic policy wordings plus endorsements, and measure how often the effective limit — not the printed one — is applied.
Developer & knowledge tools
Tessl
Structured, versioned context for coding agents — an agent enablement platform. [source]
The hard question is likely attribution: proving that a given unit of injected context actually changed agent behaviour, rather than riding along with it.
I'd run ablation evals — the same repo-level tasks with and without each context unit — and publish the per-unit lift, because that number is the product.
Granola
An AI meeting notepad that enhances your own notes — no meeting bot. [source]
I'd expect the hard part to be fusing imperfect transcription with a user's fragmentary notes — especially attributing decisions and action items to the right owner.
I'd build an eval on synthetic multi-speaker transcripts with deliberately ambiguous ownership, graded on action-item attribution precision.
UnlikelyAI
A neurosymbolic platform aiming at controllable, auditable, hallucination-free AI. [source]
The risk likely concentrates at the LLM-to-symbolic boundary: a mistranslation from natural language into the formal representation fails silently, with full apparent auditability.
I'd write a round-trip fidelity eval — natural language to formal representation and back — weighted toward ambiguous and underspecified inputs.
Science
Latent Labs
London frontier-AI lab building generative models for protein design (Latent-X, Latent-Y). [source]
The hard deployment problem is likely eval design itself: ground truth needs a wet lab, so the in-silico metrics partners see must be shown to predict lab outcomes.
I'd instrument the platform's proxy metrics against downstream validation results and build the calibration report that tells a partner how much to trust a score.