Four systems in production, and the figures each one moved — measured after go-live, from the client's own reporting, against a baseline agreed before we wrote anything. The method is on this page too, because a percentage without one is decoration.
Every figure below is attached to a named client. Nothing here is from a pilot.
Product lines we know deeply enough to move fast in — each one delivered with the tracing, evaluation and guardrails a production AI system needs from day one.
Payroll, attendance and appraisals on one tenant-isolated platform, with an agent that answers policy questions from the handbook and routes the exceptions to a named person.
From
₹18,000per month
Best for employers of 200 to 5,000 staff
Product discovery, order status and returns answered from your live catalogue and order system rather than a scripted flow, with anything it cannot confirm handed to a human.
From
₹16,000per month + usage
Best for D2C brands and online retailers
Job intake, pricing and the matching of a vetted professional to a slot, run against live availability rather than a callback queue, with the exceptions a dispatcher actually needs to see raised as they happen.
From
₹14,000per month + usage
Best for service networks and franchises
Adaptive coursework, auto-marked assessment and a tutor agent grounded in the institution's own material rather than the open web.
From
₹15,000per month
Best for institutions and training providers
01 · HRMS · Multi-tenant SaaS
Payroll, attendance and appraisals on one tenant-isolated platform, with an agent that answers policy questions from the handbook and routes the exceptions to a named person.
From
₹18,000per month
Best for employers of 200 to 5,000 staff
02 · E-commerce · Storefront agents
Product discovery, order status and returns answered from your live catalogue and order system rather than a scripted flow, with anything it cannot confirm handed to a human.
From
₹16,000per month + usage
Best for D2C brands and online retailers
03 · Home Services · Dispatch agents
Job intake, pricing and the matching of a vetted professional to a slot, run against live availability rather than a callback queue, with the exceptions a dispatcher actually needs to see raised as they happen.
From
₹14,000per month + usage
Best for service networks and franchises
04 · EdTech · Learning platforms
Adaptive coursework, auto-marked assessment and a tutor agent grounded in the institution's own material rather than the open web.
From
₹15,000per month
Best for institutions and training providers
Four of the domains we work in. The full list is longer.
Anyone can publish a percentage. What makes one worth reading is the method behind it, so ours comes before the results rather than in a footnote under them.
Not in a pilot, not on a held-out sample, and not on the week we launched. Every figure here is from a system carrying real traffic for at least a quarter.
The numbers come out of their dashboards, not ours. Where we instrumented the measurement, their team owns the query that produces it.
What "better" meant was written down before we built anything. A result measured against a baseline chosen afterwards is a result chosen afterwards.
An anonymised case study is an unverifiable one. If a client will not be named, the work does not appear here — however good the number is.
What you will not find on this page
What each one looked like before, what actually shipped, and the control that made it acceptable to the people who had to sign it off.
We replaced a 40-person exception desk with a supervised agent that watches carrier feeds, weather and customs events, then rebooks affected shipments inside its authority limit and escalates the rest with a recommended action already drafted.
Before
Forty people working a shared exception queue by hand, each rebooking decision taking a phone call and an average of 90 minutes.
What shipped
The control: Every rebooking above the authority limit routes to a named human, and the agent cannot act twice on the same shipment without one.
Every other vendor showed us a demo. Chronixa showed us an evaluation suite running against our own data, and told us which two of our five use cases were not worth building. We signed the next week.
A retrieval system over 1.2 million internal documents — protocols, assay results, regulatory correspondence — with citation enforcement strict enough that the compliance team signed off on it for regulatory submissions.
Before
Twenty years of protocols, assay results and regulatory correspondence searchable only by filename, across four systems that did not talk to each other.
What shipped
The control: No answer ships without a source. The compliance team cleared it for regulatory submissions on exactly that basis.
Our compliance team has killed three AI projects. This is the first one they approved — because the citations were enforced at generation time rather than checked afterwards. That distinction mattered enormously.
Document intelligence across KYC packs, source-of-funds evidence and corporate structures. Confidence-gated: anything below threshold routes to a human reviewer with the uncertain field already highlighted.
Before
KYC packs read by hand over four days, with corporate structures reconstructed manually from PDFs of varying quality.
What shipped
The control: Anything below threshold routes to a human reviewer with the uncertain field already highlighted — not the whole document again.
Nine weeks from kickoff to an agent handling live onboarding traffic, in our VPC, under our security model. I have run programmes with four times the budget that delivered a fraction of that.
We rebuilt the retrieval layer behind a patient support desk so answers come from the current care protocol rather than whichever wiki page ranked highest, and routed anything the index could not ground to a clinician with the source already attached.
Before
A support desk answering from a wiki nobody had pruned since 2019, where the top-ranked page for a protocol question was routinely two revisions out of date.
What shipped
The control: If the index cannot ground an answer in a current document, the agent does not attempt one — it hands the ticket to a clinician with the closest source attached.
They rebuilt our retrieval layer and our cost per resolved ticket dropped by 71%. What I did not expect was the tracing — I can now open any answer the agent gave last quarter and see exactly what it read.
Aggregated across every production engagement currently under our operations retainer. Reviewed quarterly with each client.
Executed across production deployments in the last 12 months.
Measured across all managed inference and retrieval services.
Per automated workflow, twelve months after go-live.
From signed discovery to the first agent serving live traffic.
Four things the figures above do not say. A page of evidence that will not state its own edges is a page of marketing.
These are the engagements that reached production and agreed to be named. It is a small sample, and a small sample is what a firm that turns work down looks like. Treat every figure as one system, not as an average across an industry.
The 62% is operating cost against the process it replaced. Our build and retainer are a separate line, and we will put both next to each other on a call rather than quietly nets them off here.
99.98% covers the inference and retrieval services we manage. When a model provider degrades or your source system is down, that shows in your numbers and it should — we route around it where we can and say so where we cannot.
Nothing here was measured on a public evaluation set. That makes the figures harder to compare between vendors and considerably more relevant to whether the thing works on your data.
Bring one process you think an agent could run. We'll tell you straight whether it's worth building — and what it would cost if it is.
