Training and evaluation platform for real-world voice agents: task success, compliance, timing, escalation, and conversation quality.
X For Y
Braintrust / LangSmith for voice agents
Buyers
Voice-agent startups, enterprise AI teams, contact-center AI teams, labs building agentic voice.
Opportunity
Every serious voice-agent company needs testing, regression, improvement loops, and measurable task success before deployment.
Trelis Edge
Voice expertise, multi-speaker model work, evaluation mindset, public AI credibility, and access to data-provider/RL thinking.
Risk
Customers may solve enough with prompts, logs, and frontier provider tools. Product must show obvious lift, not research novelty.
Next Test
Pick one workflow, build 50 simulated scenarios, run a customer's agent, score failures, and show improvement in one week.
ClarityHigh
UpsideHigh
FitHigh
Sell high-quality conversational datasets, eval sets, and reward data to labs and enterprise teams building voice agents.
X For Y
Scale AI for real conversational voice data
Buyers
Frontier labs, voice AI startups, robotics labs, enterprise AI teams, data/platform teams.
Opportunity
Labs and voice companies will pay for scarce, rights-clean, multi-speaker data that improves real conversation behavior.
Trelis Edge
Current multi-speaker focus, likely dataset creation path, direct ML audience, and ability to package data with evals.
Risk
Exclusive data sales can give away strategic leverage. Rights, consent, provenance, and resale terms matter.
Next Test
Pitch non-exclusive dataset/eval access to 10 labs or voice companies and ask what they would pay for.
ClarityHigh
UpsideMedium
FitHigh
AI voice support agent that answers customer calls, resolves issues, escalates cleanly, and improves from real conversations.
X For Y
Intercom / Fin for voice
Buyers
Customer support leaders, CX teams, call centers, SaaS, e-commerce, healthcare, insurance, and financial services.
Opportunity
Support calls are expensive, measurable, and repetitive. If the agent resolves real calls safely, the ROI is obvious.
Trelis Edge
Voice gym/evals, multi-speaker and interruption handling, real conversation data, and ability to optimize for task success.
Risk
Crowded with voice-agent companies. Must avoid generic phone bot positioning and pick a sharp workflow or vertical.
Next Test
Build one vertical demo: inbound support call, knowledge-base lookup, account action, compliant escalation, and scored resolution.
ClarityHigh
UpsideHigh
FitMedium
Horizontal voice infrastructure for enterprises: private, customizable, high-quality voice AI where generic models are not enough.
X For Y
Twilio / Vercel for enterprise voice agents
Buyers
Enterprise product teams, customer operations, regulated enterprises, companies adding voice to workflows.
Opportunity
Voice becomes a major interface, but enterprises need trust, privacy, domain quality, deployment control, and workflow reliability.
Trelis Edge
ASR/model work, EU/Ireland positioning, founder credibility, and a route from multi-speaker ASR into agents and data.
Risk
Kyutai/Gradium/OpenAI/Gemini are strong at the model layer. Trelis must win through product depth and workflow specificity.
Next Test
Choose one vertical workflow and sell a private voice stack pilot, not a broad "voice platform."
ClarityMedium
UpsideHigh
FitMedium
Composer-style coding harness: train, evaluate, and improve coding agents on real repository tasks using RL/evals.
X For Y
Composer / Devin training gym for enterprise codebases
Buyers
Developer-tool companies, enterprise engineering teams, AI labs, companies with large private repos.
Opportunity
Coding agents are valuable and measurable. Enterprises need repo-specific agents that can pass tests and ship safely.
Trelis Edge
Strong AI engineering audience, practical coding workflows, and possible RL/eval harness experience.
Risk
Very crowded: Cursor, Cognition, OpenAI, Anthropic, Factory, and internal enterprise tooling.
Next Test
Build a tiny harness for one repo: issue in, patch out, tests run, reward assigned. Show measurable improvement.
ClarityMedium
UpsideHigh
FitMedium
Computer-use automation for organisations: agents that perform browser/desktop workflows with monitoring and controls.
X For Y
UiPath for AI computer-use agents
Buyers
Operations teams, finance/admin teams, BPOs, regulated back-office teams, workflow-heavy SMEs.
Opportunity
Back-office work is full of repetitive browser tasks, data entry, approvals, and workflow handoffs.
Trelis Edge
General agent/tooling expertise and potential enterprise workflow angle.
Risk
Operationally messy, crowded, and not obviously connected to current Trelis voice/data edge.
Next Test
Find one painful internal workflow in a company and automate it end-to-end with audit logs.
ClarityMedium
UpsideMedium
FitLow
Sovereign AI for Ireland/EU: regulated, private, auditable AI infrastructure and models for sensitive sectors.
X For Y
Palantir / Mistral-style AI stack for regulated EU voice and agents
Buyers
Government, healthcare, finance, insurance, enterprise compliance teams, EU public-sector bodies.
Opportunity
Regulation, data residency, national capability, and enterprise trust will matter more as AI becomes critical infrastructure.
Trelis Edge
Irish/EU base, technical credibility, public ML audience, and a possible regulated voice/data wedge.
Risk
"Sovereign AI" is too abstract without a product. It needs a concrete buyer and first workload.
Next Test
Anchor it to voice: private enterprise voice agents and data/evals for regulated EU workflows.
ClarityLow
UpsideHigh
FitMedium
ChatGPT-style consumer product, free and ad-supported.
X For Y
Perplexity / ChatGPT for free ad-supported AI search and assistance
Buyers
Consumers as users; advertisers as buyers. Possibly education, ML learners, or technical users as first audience.
Opportunity
Consumer assistants have massive usage and ads can monetize large free audiences.
Trelis Edge
Existing ML audience, educational distribution, and ability to test with community.
Risk
Generic assistant competition is brutal. Ads may weaken trust. Distribution and retention are very hard.
Next Test
Only pursue if there is a highly specific audience or workflow where Trelis can win distribution cheaply.
ClarityLow
UpsideHigh
FitLow