Outbound and research pipelines
Research pipelines: a series of agent workflows that gather data at scale and filter it to what a decision needs. Prospects worth a call, competitors in a market, buying signals in filings and papers.
AI consulting & implementation
AI engineers who embed inside your company.
Our engineers join your team, learn how the work gets done, and build agents with measurable ROI.
We’ve built for
What we do
Support and knowledge assistants
Answer engines over your own documentation, code or past tickets, inside Slack, Teams or WhatsApp.
Invoice and document intake
Invoices, orders and carrier notices that arrive by email and PDF, read into the systems you already run. The fields pulled out and matched against the purchase order or the contract.
Audits of AI already running
We measure the AI you have against your real questions and records. Then we tell you what to fix.
Token cost optimization
We find where your AI spends too much on tokens and fix it: shorter prompts, cheaper models where quality holds, caching, and fewer tool calls per task. Each change is tested on a replay of your own past requests before it ships, then checked against live traffic after.
Custom builds
An agent, a pipeline or an internal tool that fits no category above, wired into your data and tools. Scoped from an audit and built in weekly increments you can see. The same rules as everything else here: a test on your real cases before launch, a weekly report after.
What we’ve built
Token cost optimizationThesis Finance
Reducing AI costs in an investing assistant
- Problem
- Thesis Finance uses AI to answer questions about portfolios and stock options. Some answers required several model calls that repeatedly charged for the same background information.
- Solution
- We added caching so the model could reuse that information at a lower cost, and removed duplicate and unnecessary data from the options results sent to it.
- Outcomes
- 57% lower model cost in an offline replay of 11 recorded options requests: $2.09 to $0.90 per answer.
- Options data sent to the model was 2.5–6× smaller. The savings have not been measured on live traffic.
Support and knowledge assistantBalancer, Solana Foundation, Binance
AI assistants for developer support in Discord
- Problem
- Developers building on blockchain platforms asked technical questions in Discord, including questions the documentation didn’t cover. Support teams had to answer them manually, often revisiting questions they had answered before.
- Solution
- We built assistants that searched each client’s documentation, code, and previous community answers to respond with sources. A dashboard let client teams review questions, correct answers, and add information the assistant could use next time.
- Outcomes
- Deployed in Balancer’s Discord community; used internally by Binance’s BNB Chain team.
- In BNB Chain’s test of 50 technical answers, the average accuracy score was 83.9%. Half scored 90% or higher.
Outbound and research systemBiobot Analytics
35 sales meetings for Biobot’s expansion into pharma
- Problem
- Biobot uses wastewater samples to track the spread of disease. It wanted to sell that data to pharmaceutical companies and needed help finding the right buyers and reaching them.
- Solution
- We built AI tools to research target companies, assess their fit, and draft messages explaining how each could use Biobot’s data. Our two-person team reviewed the research and ran the calls, emails, and LinkedIn outreach.
- Outcomes
- 35 meetings booked; 27 billed as qualified meetings with client-approved accounts.
- $2.4M in potential sales opportunities, according to our campaign records.
Outbound and research systemPrompt Security
Finding security buyers through a reseller’s LinkedIn audience
- Problem
- Prompt Security helps companies manage the security risks of using AI. It needed to reach security and AI leaders at banks and financial services firms. Direct outreach was difficult: one bank’s security chief would only consider vendors through its reseller.
- Solution
- We used AI to review the job titles of people following a security reseller on LinkedIn and select roles relevant to Prompt Security. Alongside this list, we researched target companies and pursued introductions through partners.
- Outcomes
- 4,036 potential contacts selected from 8,613 profiles, based on their job titles.
- Three weekly reports from the wider outreach campaign recorded two meetings booked and one tentative meeting.
How we work
Audit
A structured audit of how work moves through your company, across teams, systems and decision points.
We sit with the people who run it and their real records. Then we find where AI can cut overhead, speed the work up, or replace manual coordination for good.
Design and build
We design and build the system that will do the work and wire it into the tools, data and rules behind the process. You see it grow week by week. Your team reviews what it produces until it is reliable enough to run on its own.
Go-live
We put the system into production and connect it to your live tools. It starts doing real work. Any integration or migration work, the hosting environment and the model accounts are agreed with your team before launch.
Ongoing
Operate and improve
After go-live we stay on. We keep improving how the system runs: tighter decision rules, fewer failures, and the next piece of work added when it is ready.
Every week you get a short written update on what changed and what comes next.
Accuracy on a client’s own 50 questions
| Measure | Value |
|---|---|
| Questions scored | 50 |
| Rated 90% or better | 25 |
| Average score | 83.9% |
Fig. 1a client scoring the deployed assistant on its own questionsBNB Chain, August 2023. Average score 83.9%from the record
Questions operators ask first
The answers we give on the first call, written down.
Anything this leaves open is a question for the first call.
Book a callDo I need to know exactly what to build?
No. Bring a process you’d like to improve, a task that takes too much time, or an idea you want to explore. We’ll work through it with you and recommend a starting point.
Do we need a technical team?
We handle the engineering. We’ll need someone who knows the process to answer questions and review what we build, plus help from whoever manages access to your software and systems.
Can you work with our existing software?
Yes. We start by checking how your current tools store and share information. That tells us what we can connect to, what access we need, and whether any changes are required.
How long does a project take?
The schedule depends on the scope, data, access to your systems, and testing. We agree on the stages and target dates before starting, and show you progress as we build.
How do you price a project?
We agree on the audit scope and fee in advance. For a build, we put the scope, timeline, and price in a written proposal. Ongoing support and the costs of hosting, AI models, and other services are set out separately so you can budget for running the system too.
Who owns the software and data?
Your data remains yours. We set out ownership of the custom code, any third-party licenses, and access to the running system in the agreement before work starts. If your team will take over, we include the documentation and training they’ll need in the scope.
What happens to our data?
We agree on what data the system needs, where it will run, and who can access it. Your IT or security team helps us check how that data is stored and shared with AI providers. We settle this before connecting to your systems.
What happens after launch?
We monitor the system, fix issues, and improve it under an agreed support plan. You get a weekly update on results and planned changes. Before launch, we agree on how to report problems and when you can expect a response. We can also help your team take over if you want to run it yourself.
Book a call
Thirty minutes to talk through what you’d like to improve.
Book a callChoose a time that works for you.