Skip to content
Doing AI

Work

Six areas. Eight sectors. One pattern.

Intake, exceptions and follow-through, wired into systems you already run. The vocabulary changes by sector; the shape of the work does not.

What we do

Six areas where AI changes throughput.

01

Revenue operations

Respond to every enquiry in seconds and keep the pipeline moving without manual follow-up.

  • Inbound leads qualified, enriched and routed on arrival
  • Research and follow-up sequences run automatically across the pipeline
  • Proposals and quotes drafted from your templates and pricing

02

Operations

Remove the manual handling that sits between your systems.

  • Documents, invoices and data entry processed end to end
  • Systems connected directly, eliminating re-entry between tools
  • Approvals and exceptions routed by defined rules

03

Customer support

Resolve routine enquiries instantly, in your brand's language, at any hour.

  • Agents that close the majority of routine tickets without escalation
  • Grounded in your documentation, policies and ticket history
  • Structured handover with full context when a person is needed

04

Marketing and content

Produce more on-brand output at a fraction of the effort.

  • Content, ad copy and SEO assets generated from your own data
  • Personalisation across segments and campaigns
  • Continuous performance monitoring and reporting

05

Finance and back-office

Close faster, with fewer corrections and less chasing.

  • Reconciliation, reporting and compliance checks
  • Dashboards that maintain themselves
  • Alerts triggered the moment a defined threshold moves

06

Data and knowledge

Make what your organisation already knows usable on demand.

  • Natural-language access to internal knowledge, answers cited to source
  • One searchable, permissioned layer across your systems
  • Governed, auditable, and retained entirely under your control

Case studies

Systems that reached production.

Every figure below is measured, with the method stated next to it.

A global education groupEducation · admissions and fee operations · live since June 2026

The compliance queue that used to be a spreadsheet

Delivered by Nithish Singh through Adidev Industries, January–July 2026, before Doing AI existed.

Name withheld under the confidentiality clause of the governing agreement. It can be released if the client agrees in writing.

Before

Admissions ran across twenty repositories and a dozen services, and the checks that decide whether a student's documents and fees are in order lived beside them in spreadsheets and inboxes. Nobody could say what was outstanding without assembling the answer by hand, and a fee ledger that read zero silently blocked students from paying at all.

What we built

A compliance audit workflow taken from beta to production across the admissions dashboard and four backend services: a live verification queue with per-document status, round-robin assignment to named agents, an audited contact-edit path with OTP, role separation between manager and agent, and a payment receipt pipeline. It runs inside the systems the team already used — no new tool to log into.

Revenue double-counting corrected
₹3,14,000
Self-paid orders on the deactivated source system, reversibly cancelled during the migration so finance reconciled against the real records.
Share of all engineering across the platform
15.6%, ranked 1st of 43
Authored non-merge commits across 20 repositories, 1 Jan – 2 Jul 2026, identities normalised and bot commits excluded. Reproducible from source control.
Production deployments in the heaviest month
43
Automated image updates through CI to the production cluster during June 2026.

How this was measured: Every figure is taken from monthly engagement reports prepared for the client's finance and management review, each derived from source control and production records rather than estimated afterwards.

How you can check it: The contribution figures are reproducible from git history across the twenty repositories, with author identities normalised and bot commits excluded. The report they come from was delivered to the client's own finance and management for review, not written afterwards for a website. We will walk a prospective client through it under NDA.

A global education groupEducation · finance and accounting operations · live since July 2026

The ledger balanced until somebody actually checked

Delivered by Nithish Singh through Adidev Industries, July 2026, before Doing AI existed.

Name withheld under the confidentiality clause of the governing agreement. The defects below are described by kind and scale; the underlying amounts are the client's own figures.

Before

The accounting export that fed the finance system ran every month without errors, so nobody had reason to doubt it. It was wrong in five separate ways at once. A whole class of recovered payments was being discarded by a filter that named the payment types it accepted; admission vouchers were missing from the journal entirely; one collection was booked twice with both legs claiming the same payment; and the export was unscoped, so other institutions leaked into a single institution's books.

What we built

The export was rebuilt and then corrected against real accounting data — each defect found by reconciling the journal rather than by a test failing. The payment-type whitelist was removed rather than extended, because adding the missing type would only have guaranteed that the next unanticipated one was dropped just as silently. Revenue is now keyed by intake cohort rather than payment type, vouchers clear at the money that actually arrived, and a loan counts as a loan only once the lender has disbursed.

Recovered payments silently discarded
304 of 304
A payment class matching neither branch of an accepted-types filter, dropped in full — a 100% loss with no error raised and no failing test.
Student fee ledgers showing a wrong paid amount
~28,000
Roughly 95% of all paid ledgers, where the paid column was double-counted platform-wide. Fixed and shipped 31 July 2026.
Distinct reporting defects found by reconciliation
5
Each identified by reconciling the general ledger against source records, not by an alert, an exception or a failing test. None had ever surfaced as an error.
Student records migrated, each ledger-verified
158
Four migration operations, every one dry-run before commit and checked against the live ledger rather than a database read-back. No money moved.

How this was measured: Every defect above was found by reconciling a general ledger against the systems that produced it, and every figure comes from the monthly report prepared for the client's own finance review. The money amounts involved are the client's and are not published here.

How you can check it: Each defect is recorded against the commit that fixed it and the production release it shipped in. The client's finance team reviewed and accepted the corrections — the strongest verification available for accounting work, because the party who would object is the party who signed it off.

A global education groupEducation · student communications · live since April 2026

Adding one WhatsApp message used to mean a deploy

Delivered by Nithish Singh through Adidev Industries, April 2026, before Doing AI existed.

Name withheld under the confidentiality clause of the governing agreement. It can be released if the client agrees in writing.

Before

Transactional WhatsApp messages were hand-coded per institution or pushed through a legacy mailer with no campaign routing. There was no kill switch, no way to build the parameters for a new event, and adding one meant editing the binary and redeploying. Three institutions each needed a different message shape, so each new message multiplied.

What we built

One communications backbone across WhatsApp and email, driven by configuration rather than code. Nine student lifecycle events wired through per-event parameter builders, per-institution template shapes, a kill switch, and defaults moved into config so a new campaign works in production without a release.

Institutions live on one shared pipeline
3
Each with its own template-parameter shape, previously requiring separate hand-written handling.
Student lifecycle events wired
9 + welcome + document rejection
Enquiry, interview cleared and not cleared, fee verified, and the registration and course fee events, each with its own parameter builder.
Redeploys needed to launch a new campaign
None
Defaults merged into configuration, so campaigns are added without touching or shipping the binary. Previously every new event required a code change and a release.

How this was measured: Taken from the April 2026 engagement report, which records each change against the service it shipped in. The pipeline built here is what allowed a new institution to launch the following month with no per-event work at all.

How you can check it: The claim that a new campaign needs no release is checkable the plain way: the next institution went live the following month without a single per-event code change, which is recorded in the May report.

A global education groupEducation · internal operations · live since May 2026

Four internal tools became one platform

Built by Nithish Singh at Adidev Industries, May–July 2026, before Doing AI existed. The platform is Adidev's own work rather than a change to the client's systems.

Name withheld under the confidentiality clause of the governing agreement. It can be released if the client agrees in writing.

Before

The path from a sales enquiry to an admitted student ran through four separate internal applications — a sales intelligence tool, a reporting tool, a set of dashboards, and an internal app — none of which agreed with the others. Finance reconciled bank statements to receipts by hand, in spreadsheets.

What we built

One platform carrying the whole funnel: a governed data warehouse with dimensional modelling, a bank-statement reconciliation engine with rule-based matching and a manual workbench for the rest, scheduled reporting, and operator surfaces for the teams that use it. Role and row-level security throughout, with an append-only audit log.

Internal applications consolidated
4 → 1
A sales intelligence tool, a reporting engine, dashboards and an internal app, replaced by a single platform on its own cluster.
Access control model
79 permissions across 26 resources
A resource-by-action permission system with ten system roles and row-level security, replacing per-application access rules.
Manual finance reconciliation
Replaced by a workbench
Bank-statement to receipt matching moved from spreadsheets and hand-matching to rule-based ingestion with exact-reference matching, delete-with-undo, and suggestions for the remainder. A recurring monthly task, not a one-off.
Production deployments in one month
28
Date-tagged image deploys during July 2026, on a platform sustained across all 31 days of that month.

How this was measured: Figures come from the May, June and July 2026 engagement reports. Commit counts are deliberately left out: they measure activity rather than value, as those reports themselves note.

How you can check it: The platform runs in its own namespace with date-tagged image deploys, so its release history is a matter of record rather than recollection. The permission model can be read straight out of the schema.

BillGill — our own productSales operations · any sector that runs a phone team · live since 2026

The call data existed. Nobody could act on it.

Built and operated by Adidev Industries. Live.

Before

Call activity lived in two separate telephony systems that did not talk to each other, and neither connected to the sales pipeline. Answering the only question that mattered — is calling activity actually producing enrolments — meant exporting from both, matching by hand in a spreadsheet, and accepting that the answer was already out of date.

What we built

A call analytics platform that ingests both telephony sources into one place, joins them to pipeline outcomes, and reports employee productivity, lead status and sales performance from a single view. A key-gated public API exposes the same data to other systems, so the numbers stop being retyped.

Telephony systems unified
2 → 1 view
Two call platforms previously reconciled by hand, ingested into one schema with sentiment and scorecard reporting on top.
Calling joined to outcomes
Closed the loop
The platform reads the enrolment warehouse directly, so calling activity and what it produced are answerable in one query rather than two exports.
Integration surface
Key-gated public API
Documented read-only endpoints with case-insensitive filtering, so downstream systems consume the data rather than re-entering it.

How this was measured: Described from the product itself and from the monthly reports that record its development and its security-audit pass.

How you can check it: It is our product, so the whole of it is available to look at — the running system, the API and its history. Nothing here depends on a client confirming anything.

SolveX — our own productReporting and finance operations · any multi-system business · live since 2026

Eleven reports that used to be assembled by hand

Built and operated by Adidev Industries. Live.

Before

The numbers that finance and operations needed were spread across five separate databases and two payment gateways. Every recurring report meant pulling from each, reconciling the differences, and rebuilding the same spreadsheet — every period, by a person, with the errors that implies.

What we built

A reporting pipeline that pulls all seven sources into one warehouse and generates the reports directly from it: eleven report types across four institutions, delivered as spreadsheets by email or self-service, on a schedule. Loan and lender reconciliation, per-order renewal detection and settlement tracking are computed rather than assembled.

Sources consolidated
7 → 1 warehouse
Five operational databases and two payment-gateway APIs, previously exported and reconciled separately for each report.
Report types generated automatically
11
Across four institutions, delivered by scheduled email and self-service download instead of being compiled and sent by hand.
Reporting effort per period
Recurring, not one-off
The saving repeats every reporting period rather than being a single migration — which is the difference between a project and an asset.

How this was measured: Described from the product and from the monthly reports covering its build and its cross-cloud repointing onto managed database clusters.

How you can check it: Ours to show in full. The report definitions, the warehouse schema and the scheduling are all inspectable, and the outputs are the ones finance actually uses.

A global education groupReporting integrity · any business that exports data · live since April 2026

The export quietly returned ten rows

Found and fixed by Nithish Singh at Adidev Industries, April 2026.

Name withheld under the confidentiality clause of the governing agreement.

Before

A listing endpoint that fed management reports had a default: when no row limit was passed, it returned ten. Not an error, not a warning, not a truncation notice — ten rows, correctly formatted, indistinguishable from a complete answer. Every report built on it had been quietly wrong for as long as it had existed, and nothing anywhere would ever have said so.

What we built

The silent default was removed and replaced with an explicit ceiling high enough to return the real answer. The wider lesson was applied where it belonged: a query that cannot return everything should say so, rather than returning a plausible fraction.

Rows a full export actually returned
10
The unlimited case fell through to a ten-row default. Reports were built on it and read as complete.
How it announced itself
It did not
No error, no exception, no failing test. The bug was found by someone reading the code path, not by anything the system reported.
Class of defect closed
Silent truncation
Fixed at the default rather than at each call site, so a future caller that forgets a limit gets the answer instead of a fraction of it.

How this was measured: Recorded in the April 2026 engagement report against the commit and service that fixed it.

How you can check it: This is the kind of defect worth judging us on, because it is the kind nobody is paid to find. Ask us on a call what we would look for first in your reporting, and see whether the answer sounds like someone who has been here before.

A regulated institutionRegulated review · lending, claims, onboarding, compliance · live since May 2026

Three pairs of eyes, and a record of whose

Delivered by Nithish Singh through Adidev Industries, May 2026.

Name withheld under the confidentiality clause of the governing agreement.

Before

Approvals that regulation required to pass three separate reviewers were tracked as a single yes or no. Who checked what, in which order, and on what basis was reconstructed afterwards from memory and email — which is the same as not having it, on the day somebody asks.

What we built

A configurable multi-tier review pipeline: three reviewer levels with distinct roles and permissions, per-document routing between them, and an audit trail written as the work happens rather than assembled afterwards. Institutions with different approval rules configure the pipeline instead of getting a fork of it.

Review tiers with separate roles
3
Each tier a distinct role with its own permissions, so a reviewer cannot complete a stage that is not theirs.
Audit trail
Written as it happens
Captured at the moment of each decision rather than reconstructed later, which is the only version that survives being questioned.
Configurable per institution
Yes
Approval rules are configuration, not a code branch — so a second institution with different requirements does not become a second codebase to maintain.

How this was measured: Taken from the May 2026 engagement report, which records the role model and the tier that completed it.

How you can check it: The permission model and the audit schema are the artefact — they can be read directly and either separate duties properly or do not. The institution it was built for is regulated, and accepted it.

A global education groupBilling and pricing · any business with discounts or instalments · live since June 2026

Nobody could agree what was actually owed

Delivered by Nithish Singh through Adidev Industries, March–June 2026.

Name withheld under the confidentiality clause of the governing agreement.

Before

What a customer owed was computed in more than one place, and the places disagreed. Discounts were applied against a price that could change after the fact, so historic orders stopped resolving. Some charges were excluded from the calculation entirely, leaving balances that read as nothing owed. Finance, the operations team and the customer's own screen could each produce a different number, and each could defend it.

What we built

One calculation, used everywhere. Historic prices are pinned so a past order still resolves against the price it was actually sold at, every charge type is included in the balance, instalments carry a running per-period balance, and the amount paid is stored as the amount paid rather than derived from a discount.

Balances reading zero when money was owed
~11,000
Applications whose ledger showed nothing payable because a charge type was excluded from the calculation, so the customer could not pay even if they wanted to.
Historic orders that stopped resolving
Fixed at the source
The first price and discount row of each chain was backdated, so orders paid under an earlier price resolve correctly instead of failing.
Places the amount owed is now computed
One
Previously several, each defensible and none authoritative. Disagreement about the number was the defect, not any single wrong answer.

How this was measured: Drawn from the March to June 2026 engagement reports, each fix recorded against the release it shipped in.

How you can check it: The corrections were accepted by the finance team who had been living with the discrepancy, which is the only sign-off that means anything on a billing question.

Adidev Industries — our own operationsDelivery reliability · any team shipping to production · live since July 2026

Nine portals that could not ship themselves

Diagnosed and fixed by Nithish Singh at Adidev Industries, July 2026.

Before

The automation that promotes a built image into production had stopped being able to write back, so every release was being nudged along by hand. Nothing was broken loudly enough to raise an alarm; it had simply become normal for a person to do the last step, and nine portals were frozen behind it.

What we built

The cause was traced to an access token holding a permission level too low to write, rather than to the deployment tooling everyone suspected. Fixed at the credential, the automation resumed, and the portals unfroze.

Portals unblocked
9
Each frozen behind the same write-back failure, all released once the underlying permission was corrected.
Root cause
A token, not the tooling
A guest-level access token that could read but not write back. The deployment system, which had absorbed the blame, was working correctly throughout.
Manual steps per release
Back to zero
Hand-bumping had quietly become the routine. The fix removed the step rather than documenting it.

How this was measured: Recorded in the July 2026 report as a production infrastructure blocker, alongside the releases it unblocked.

How you can check it: Our own infrastructure, so there is nothing to take on trust — the deployment history shows the gap and shows it closing.

Your world

See what it does in your business.

purchase orders · deliveries · dispatch · invoices

Procurement

Supplier follow-up that runs itself

Delivery confirmations are requested, read and recorded against the purchase order — including the ones that arrive as a photo on WhatsApp — and only the silent suppliers escalate.

manual follow-upescalations only

Accounts

The invoice, the order and the delivery agree by themselves

Every invoice is checked against what was ordered and what actually arrived. Where the three disagree, purchasing hears about it the same day, with the likely cause attached.

found at month endflagged on arrival

Shop floor

Paperwork off the floor

Delivery notes, receipts and invoices are read as they arrive, matched to the order, and written into the system your team already uses — without anyone typing them again.

re-typingmatched and filed

What changes

What this tends to move.

Less re-entry

Connect the records your team currently copies by hand

Clear exceptions

Route discrepancies with the evidence needed to review them

Human approval

Keep consequential decisions with your team

Measured impact

Compare handling time and accuracy against an agreed baseline

We measure the starting point before agreeing targets. Savings and delivery dates depend on the workflow, data quality and integrations in your written scope.

  • A global education group
  • A commercial-vehicle manufacturer
  • A multi-site logistics operator
  • A private hospital group
  • An ed-tech platform
  • A certification body