Custom AI that does a real job
Most AI pitches are demos. Ours is a system we run our own business on: sourcing, screening and assessing candidates every working day, at a volume no manual team could match. We build the same kind of thing for other businesses.
We run AI in production, on our own P&L
Not a pilot, not a demo built for a pitch.
Our recruitment arm runs on AI we built. It reads every profile that comes in rather than the first fifty, scores each one against the criteria agreed at intake, and writes out its reasoning including the gaps worth asking about. It searches on meaning rather than keywords, so it surfaces people whose experience fits even when their CV never uses the client's terminology.
Around that sit the smaller pieces that make it usable day to day: job descriptions drafted from a scorecard, structured interview support, candidate fitment evaluation against how a specific team actually operates, search across LinkedIn and Naukri, and a Chrome extension that screens a CV in the browser where a recruiter is already working.
None of that is a case study we commissioned. It is the operational software of a business with revenue attached, which means we have had to live with every decision in it. That is a materially different experience from shipping an AI project and moving on.
Four things AI is genuinely good at
We would rather be specific about where it works than sell you a category.
Reading documents
- CVs, invoices, forms, reports
- Extracting fields reliably
- Scoring against your criteria
- Written reasoning, not just a number
Automating workflow steps
- Classifying and routing work
- Drafting from a template and context
- Flagging exceptions for a human
- Removing the copy and paste
Internal assistants
- Questions answered from your own documents
- Grounded in your data, not the web
- Access controlled by role
- Answers that cite where they came from
Browser and workflow extensions
- Working where your team already works
- No new system to log into
- Screening and summarising in place
- Adopted because it saves time
The model is the easy part
Getting a language model to produce something plausible takes an afternoon. Getting a system you can run a business on takes considerably longer, and almost none of the difficulty is in the model itself.
It is in deciding what the thing is actually allowed to conclude. What happens when it is unsure. Whether a person sees the output before it matters, and whether that person is given enough reasoning to disagree usefully. Whether the output is consistent enough that two similar inputs on different days get treated the same way.
We learned that by running one. Our own system ranks and narrows, a recruiter reviews, and nobody is rejected by software alone. Where the AI and the human disagree, we surface it, because that disagreement is usually the most informative signal in the whole process. We build client systems on the same principle.
We will tell you when AI is the wrong answer
If a rule would do the job
Write the rule. It is cheaper, faster, and it will never surprise you at the worst possible moment.
If the process is not agreed
AI applied to a process nobody has settled just automates the disagreement. Fix the process first.
If nobody will check the output
Automation nobody reviews is a risk with a schedule attached. If there is no capacity to check it, we would rather not build it.
See also: Custom software development in Pune · Internal tools · What we have built
Questions we get asked
Is this real AI or a wrapper around ChatGPT?
Both descriptions miss the point. Most useful business AI does involve a language model, ours included. What decides whether it works is everything around the model: what you feed it, how you constrain it, what happens when it is unsure, and whether a person checks the output before it matters. That is the engineering, and it is what we have spent our own time on.
What if the AI gets something wrong?
It will, and any firm telling you otherwise has not run one in production. The question is what happens next. In our own system nothing is rejected by software alone, every score carries written reasoning a person can argue with, and disagreements between the AI and the human are surfaced rather than hidden. We build client systems the same way. AI that cannot show its working is not usable for decisions that matter.
Do we need a lot of data to start?
Usually less than people assume. Reading documents, extracting fields, classifying and routing work do not require you to have a large historical dataset. Where you would need real volume is training something bespoke on your own patterns, and for most small and mid-size businesses that is not the useful starting point.
Where does AI genuinely not help?
Where the rules are already clear and consistent. If a decision can be written as a rule, write it as a rule. It will be cheaper, faster, and it will never surprise you. AI earns its place on messy inputs and judgement heavy work, not on arithmetic.
Can it run on our own systems?
Depends what the work is and how sensitive the data is. Some of it can run against your own infrastructure, some depends on external model providers. We will tell you plainly which parts sit where before you commit, because for some businesses that answer is the deciding factor.
Tell us what needs deciding, sorted or read
Describe the work that is eating people's time. We will come back within one working day with an honest view, including whether AI is the right tool for it at all.