Generative AI
Generative AI that drafts the work and proves it right
You get generative AI doing a named job inside your product: writing the first draft, reading the document, turning a request into a structured action. Checked before it ships, watched after.
- 4.8 on Clutch across 22 reviews
- PremiumAudit.io: audit cycle cut 75%, calculation errors cut 95%
- In production for SmartREI, Tankaware, ZyloBase, KiwiWrite
Generative models are good at producing something plausible and indifferent to whether it is correct. That gap is the whole engineering problem, and a better prompt does not close it. Three things do. Every output gets checked against cases your experts graded. A run can be replayed months later. The actions the model may not take alone are written down before launch. We build that scaffolding first, because a draft nobody trusts is slower than no draft.
What this service is, and which of ours it is not
Generative AI produces language. Text, structured output, a summary, a draft, a classification written in words. If the thing you want back is prose or a filled-in form, this is the page.
Three neighbouring services get confused with it, so here is the line. Model engineering trains a model on your operational data to output a number or a label: a failure prediction, a risk score, a defect flag. RAG development is generative AI with one specific job, answering from your own document library with a citation attached. AI agent development is generative AI given permission to act, calling tools and changing records in your systems.
Most real builds cross those lines, and you get them scoped together rather than sold as four projects. The distinction matters for one reason. It decides what proof your system needs before anyone trusts it.
- 01Want a number or a label? That is model engineering, and it is a different build.
- 02Want an answer with the source passage attached? Start at RAG development.
- 03Want the model to press the button as well as write the sentence? That is an agent, and the safety work doubles.
What is running now
Counted, not estimated
22
production systems we have shipped and can name
9
of them with AI doing real work inside
50+
ESG topics Greenlight's platform scores per company
38+
countries running Broollie's meeting platform
Generative features we have actually shipped
Named systems, running for named clients.
Turning a meeting into work that gets done
Broollie drafts the agenda, assists during the call, writes the minutes, then chases every action item. Universities, government and healthcare teams run it across more than 38 countries.
Reading documents nobody wants to read
PremiumAudit.io runs the insurance premium audit lifecycle on Claude, with field mapping, validation and exception handling. A human auditor stays in the loop on every policy.
Plain words into a structured search
Military Cruise Deals lets a traveller describe the trip they want. The module asks a few follow-up questions, then converts the request into a validated search across ships, ports and dates.
Evidence weighed into a written assessment
Greenlight reads laws, certifications and controversies rather than a company's own reporting. It covers more than 50 ESG topics with 10 to 15 independent sources behind each one, and an expert verifies the result.
- ScopeFree call. Which output, judged by whom?
- GradeReal cases, answered by your experts first.
- GroundPoint it at your data, not the model's memory.
- GateName what a person must approve.
- ScoreAgainst those cases, before every release.
- RunTraces kept, drift watched, migrations planned.
Grading comes before any model work. Cases graded by your own people are the only thing that can later prove a prompt change made the system worse rather than better.
Deciding what a good answer even is
With ordinary software, correct is obvious. The invoice totals or it does not. With generated text there are many acceptable answers and many more that look acceptable and are not.
So the first workshop is not about models. We sit with the people who do the job today and collect real cases, including the awkward ones they argue about. They write the answer they would accept. Every future change gets scored against that set. It is why we can tell you in month six whether the system is still right.
- Your experts grade the cases, not us. A vendor grading its own homework proves nothing.
- The system is allowed to say it does not know, which most demos are not.
- Anything irreversible waits for a person to approve it.
- The model name lives in configuration, so a retirement means a config change and a re-run of the test set.
What we build on
Models
- Claude API
- OpenAI GPT-4o
- Perplexity Sonar-Pro
- Fireworks AI
- Deepgram Voice Agent
Build
- Python
- FastAPI
- LangGraph
- TypeScript
- React
- Celery + Redis
- PostgreSQL
Run
- AWS
- Docker
- Sentry
- GitLab CI
- GitHub Actions
Shipped integrations
- Twilio Voice
- Twilio WhatsApp
- Zoom API
- Webex API
- AWS S3
- SendGrid
- Stripe
“I am extremely happy with the results and would highly recommend Hashlogics to anyone.”
Daniel Khin · CEO, PremiumAudit.io
The usual generative AI project against ours
Both look identical in a demo. They separate at the point where somebody has to act on the output without reading it twice.
What counts as correct
The usual approach
Whatever looked good in the demo.
How we build
Written down by your experts before we build.
Proof it still works
The usual approach
Nobody re-checks after launch.
How we build
The graded set re-runs on every change.
When it is wrong
The usual approach
A customer finds it first.
How we build
Replay the run: input, retrieval, tool calls, output.
Irreversible actions
The usual approach
The model does them and logs a success.
How we build
Blocked. A named person approves, and that is in the design.
Model retirement
The usual approach
Discovered when the API starts erroring.
How we build
Planned. Model id in config, test set ready to prove the replacement.
After launch
The usual approach
The contract ends at go-live.
How we build
An agreed service level, or a documented handover to your team.
01How do we stop it inventing things in front of a customer?+
You ground it and you gate it. Grounding means the system answers from your documents and cites the passage it used, so a reviewer can check in seconds. Gating means the actions you cannot take back wait for a person. Where your documents do not cover a question, the system says so instead of guessing. Most demos are missing that behaviour.
02Which model should we build on, and are we locked in?+
We pick per job and keep the model name in configuration, so switching is a config change plus a re-run of your graded cases. We ship on Claude, OpenAI and Fireworks depending on the task and the data rules. Anthropic gives at least 60 days' notice before retiring a public model. That is plenty of time if your test set exists, and nowhere near enough if it does not.
03Can generative AI go into a product that already exists?+
Yes, and most of ours did. Military Cruise Deals got an AI search module inside a WordPress site they already ran. Scoping conversations cost nothing. Every build starts with a fixed-fee Blueprint: an engineer reads your real systems and writes the plan, with a fixed price for each milestone rather than a guess. The fee is credited in full against the build, and if the plan isn't one you'd act on, you don't pay for it.
04What does it cost to run every month?+
Model spend tracks volume and how much text each call carries, and it is usually small next to the engineering around it. An internal tool answering a few hundred queries a day costs little. The lever that actually moves the bill is how much context you send on every call, which is a design decision we make with you.
05Do you fine-tune models?+
Rarely, and we will tell you when it is not the answer. Most problems people bring us as fine-tuning turn out to be retrieval problems or unclear success criteria. Fine-tuning changes how a model writes; it does not reliably teach it new facts, and it makes every future model migration harder.
06Who owns what you build?+
You do. Source code, prompts, the graded test cases, the data pipelines and the infrastructure accounts are yours, in writing, before work starts. The eval set matters most here, because it is the asset that lets any competent team take over from us later. Every build also includes the first 2 months of support and maintenance, free.
07We are a startup. Is this only worth it at enterprise scale?+
One narrow feature shipped properly is the normal way to start, and several systems we run began that way. Golancer is a small team's product with AI-generated daily priorities inside it. The scaffolding scales down: a small graded set still beats no graded set.
08Our compliance team will ask where the data goes. What do we tell them?+
The honest answer names your provider and your plan, and we confirm it in writing during scoping rather than from memory. We build for data rules rather than around them. TrialTriage runs on de-identified patient data with field-level masking and an audit trail across 23 tracked actions. Your compliance team gets that design documented before anything ships.
Related
- RAG development →When the answer has to come from your documents, with a citation.
- AI agent development →When the model has to act, not only write.
- hire generative AI engineers →The engineers who build this, embedded on your team.
- How to run evals on an AI agent before launch →The graded set, in practice.
- Why do AI projects fail →The failure modes this page is built against.
