Telemedicine platforms and the realities of remote care
Video consults are routine in Indian healthcare now, but the platforms that work are built around bandwidth, regulation and continuity, not around the call.
Continue readingInsights
Engineering notes and field reports
Things we learned the expensive way, written down so the next team does not have to — 80 notes and counting.
Healthcare
10 notesClinical workflow software, FHIR, ABDM and real interoperability
Health records still move by PDF and WhatsApp in most Indian care settings. What FHIR and ABDM actually change, and what they leave to implementers.
Continue readingConsent is scoped, timed, and revocable
A consent checkbox records that someone agreed once. It cannot tell you what they agreed to, for how long, or whether they have since changed their mind.
Continue readingCutting appointment no-shows with better scheduling software
No-show rates of 20 to 30 percent are common in outpatient care. Most of that gap is a scheduling design problem rather than a patient behaviour problem.
Continue readingAccess logs nobody reads are not a control
Writing every record view to a table satisfies the requirement. Detecting the receptionist who looked up a celebrity requires something to actually look.
Continue readingClinical records are append-only, including the mistakes
A note corrected in place destroys the record of what the clinician actually saw when they made the decision. That record is the point.
Continue reading
Show 4 more in Healthcare
Patient data under the DPDP Act: what clinics need to change
India's Digital Personal Data Protection Act reshapes how healthcare providers collect, store and share patient records. A practical view of the obligations.
Continue readingClinic scheduling is a constraint problem wearing a calendar
A booking needs a clinician, a room, equipment, and sometimes an interpreter — all free at once. Checking one and hoping is how double-bookings happen.
Continue readingClaims fail on data you already had
Most rejections are not clinical disputes. They are a missing field, a stale policy number, or a code that expired — all knowable before submission.
Continue readingDe-identification is harder than dropping the name column
Postcode, date of birth, and sex identify most people uniquely. Removing the obvious fields produces data that feels anonymous and is not.
Continue reading
Hotel Management
10 notesAutomating the guest journey without losing the guest
Digital check-in and automated messaging pay off in hotels, but only where the automation knows when to hand the guest to a human.
Continue readingDynamic pricing without guessing at demand
Revenue management for independent hotels is less about clever algorithms than about having pickup, pace and competitor data in one place.
Continue readingOverbooking guards that actually hold
Checking availability and then writing the booking is two statements. Under concurrency that gap is where the double-booking lives.
Continue readingChannel sync that reconciles instead of drifting
Pushing availability to the OTAs is the easy half. Knowing when your number and theirs disagree is the half that prevents walk-outs.
Continue readingThe direct booking engine is a conversion problem
Hotels spend heavily to pull traffic away from OTAs and then hand it to a booking engine that loses most of it before payment.
Continue readingFolios that split without a spreadsheet
A guest bill is not a total. It is a set of postings with owners, and the moment two people pay for parts of one stay, a single amount column stops working.
Continue reading
Show 4 more in Hotel Management
Housekeeping status is a real-time problem disguised as a CRUD screen
A room marked clean twenty minutes ago is not information. During checkout rush the front desk needs the state now, not the state at last refresh.
Continue readingChannel managers, rate parity and the cost of drift
A channel manager that is 30 seconds behind is not a sync problem, it is an overbooking and rate parity problem with a real cost attached.
Continue readingRate plans are rules, not rows
Storing a price per room per night works until the first minimum-stay promotion, and then you are writing a rules engine in the pricing query.
Continue readingThe night audit should be boring
If closing the day requires someone to be awake, click through a wizard, and hope nothing was mid-transaction, the close is not a process. It is a ritual.
Continue reading
Audit Management
10 notesLeast privilege survives only if you review it
Access granted correctly on day one becomes over-permissioned within a year. The review process, and the trail behind it, is what holds the line.
Continue readingSBOMs and the dependencies you did not choose
A modern application ships hundreds of packages nobody selected deliberately. Supply chain risk starts with knowing what is actually in the build.
Continue readingEvidence needs a home before the audit, not during it
Six weeks of screenshotting, chasing colleagues, and reconstructing last March is not audit preparation. It is a data collection project run under a deadline.
Continue readingAppend-only is what makes a trail a trail
An audit log with an UPDATE statement anywhere near it is a report. The value comes entirely from the fact that nothing can be changed after the fact.
Continue readingZero trust and the evidence it should produce
A zero trust rollout that cannot show who accessed what, under which policy, has bought controls without buying assurance.
Continue readingMap controls once, report against many frameworks
SOC 2, ISO 27001, and a customer questionnaire ask for the same evidence in different words. Collecting it three times is how compliance eats a quarter.
Continue reading
Show 4 more in Audit Management
What DPDP compliance actually requires from engineering
India's data protection law turns policy commitments into database schema, retention jobs and deletion paths that have to work on demand.
Continue readingFindings need an owner and a date, or they are just observations
A remediation tracker where everything is open, unassigned, and undated is a list of things you have written down rather than a list of things being fixed.
Continue readingAccess reviews that are not rubber stamps
Sending a manager a list of forty permissions they do not understand produces a signature, not a review. The control passes and the access stays wrong.
Continue readingContinuous monitoring beats the annual snapshot
Testing a control once a year tells you it worked on one day. What you want to know is whether it worked on all the other days.
Continue reading
Voice AI
10 notesA voice agent without write access is an FAQ
The difference between a demo and a deployment is whether the agent can safely change something in your CRM while the caller is still on the line.
Continue readingDeflection is a vanity metric on its own
A voice agent that handles 70 percent of calls without a human may be succeeding or may be hanging up on people. The measurement has to say which.
Continue readingBarge-in is not a feature, it is table stakes
An agent that keeps talking after the caller starts speaking is not a slow agent. It is a rude one, and callers hang up on rude.
Continue readingYour voice agent is only as good as the trunk
Model quality gets the attention, but most production voice AI failures trace back to telephony: codecs, jitter, DTMF, and call routing nobody tested.
Continue readingEndpointing decides how the agent feels
Deciding the caller has finished speaking is a guess made in under a second. Guess early and you interrupt; guess late and the pause feels like a dropped line.
Continue readingCode-switching is the default, not the edge case
Indian callers move between Hindi, English and a regional language inside one sentence. An agent built around a single language choice will fail most of them.
Continue reading
Show 4 more in Voice AI
Telephony audio is 8kHz and your model was not trained on it
A speech pipeline validated on studio recordings meets narrowband, packet loss, and codec artefacts on the first real call.
Continue readingA transfer that loses context is worse than no agent
The caller has already explained themselves once. If the human who picks up starts with "how can I help", the agent has cost you the call.
Continue readingIn voice AI, latency is the product
Callers judge a phone agent in the first second. Model quality barely registers if the pause before it speaks is too long.
Continue readingRecording, consent, and the transcript you will need later
Voice systems produce the most sensitive data you hold and your most useful debugging artefact. Those two facts are in tension, and both are non-negotiable.
Continue reading
AI Agents
10 notesWrite the evals before you write the prompt
Without a scored test set, every prompt change is judged by whether the last three examples looked better. That is not iteration, it is drift.
Continue readingGive the agent its own credentials, scoped down
An agent running with the service account can do anything the service can. The blast radius of a bad tool call is then the whole system.
Continue readingRAG fails at retrieval, not at generation
When the answer is wrong, the instinct is to blame the model. Nine times out of ten the right passage was never in the context to begin with.
Continue readingPut a ceiling on the loop before you ship it
An agent that can call tools can call them forever. Without a hard budget, one malformed input becomes an overnight bill and a rate-limit ban.
Continue readingMultimodal models and the business document backlog
Models that read layout, tables and photographs together are changing document processing, but only where the pipeline is built for verification.
Continue readingDesign the handoff before the happy path
Every agent hands over to a person eventually. If the transition loses context, the customer repeats themselves and the agent has cost goodwill, not saved it.
Continue reading
Show 4 more in AI Agents
How engineering teams actually use AI coding assistants
Adoption is near universal and trust is not. The teams getting real leverage have changed their review and testing habits, not just their editors.
Continue readingThe harness is the agent
A demo agent calls one API. A production one needs scoped credentials, cost ceilings, evals, and someone to escalate to. That scaffolding is most of the work.
Continue readingRAG and the enterprise knowledge problem: retrieval is the hard part
Most retrieval-augmented systems fail because the right document never reaches the model, not because the model reasoned poorly over it.
Continue readingAgentic workflows in real businesses: where they hold and where they break
Agents that do work rather than answer questions need budgets, retries and human checkpoints before they earn a place in an operational process.
Continue reading
Fintech
10 notesUse double-entry from day one, even for a prototype
The argument against it is always "we'll add it later". Later means a migration under load, with real money already wrong in production.
Continue readingIdempotency keys are not an optimisation
Every payment integration eventually double-charges someone. The fix is not better retry logic, it is making the second attempt provably free.
Continue readingMoney belongs in integers
Floating point is fine for physics and wrong for currency. The bugs it causes are small, constant, and impossible to explain to an auditor.
Continue readingReconciliation is a product feature, not a back-office chore
If finding out you disagree with the bank needs a spreadsheet and a bad afternoon, you have not finished building the payment system.
Continue readingWebhooks need a log, not a handler
Processing a webhook inline means a provider retry, a deploy, or a slow query can silently cost you an event you cannot get back.
Continue readingSettlement files are where the truth lives
Your database says the order was paid. The settlement file says what actually reached your bank account. Reconciliation is the work of closing that gap daily.
Continue reading
Show 4 more in Fintech
KYC is a state machine, not a boolean
A verified flag cannot express "documents submitted, one rejected, awaiting re-upload, provisionally allowed to transact up to a limit".
Continue readingOnboarding drop-off is a compliance design choice
Regulated onboarding does not have to lose 60 percent of applicants. Most of that loss comes from sequencing decisions, not from the rules themselves.
Continue readingCredit decisions need a replay button
If you cannot reproduce why a loan was approved eight months ago, you do not have an underwriting system, you have a scoring script.
Continue readingUPI checkout reliability is a timeout problem
Most failed UPI payments are not declines. They are collect requests that expired, callbacks that never arrived, and status checks nobody made.
Continue reading
Inventory
10 notesStock is a ledger, not a number
A quantity column tells you what you think you have. A movement log tells you how you got there, which is the only thing that helps when it is wrong.
Continue readingBatch and expiry are not attributes of a product
The moment two deliveries of the same SKU have different expiry dates, stock stops being one number per product and becomes one per batch.
Continue readingAdding a second warehouse changes every query you have
Multi-location is not a column. It is a different question — stock is no longer "how many" but "how many, where, and can it get there in time".
Continue readingBarcode scanning is a latency problem, not a hardware problem
A scanner types faster than most forms can handle. If the round trip is 400ms, a picker scanning fifty items an hour is waiting twenty seconds for nothing.
Continue readingHeadless commerce and the catalogue data problem
Decoupling the storefront is the easy part. The harder work is getting product and inventory data clean enough to serve every channel from one API.
Continue readingCycle counts beat the annual stocktake
Closing the warehouse for two days once a year finds every error at the point it is oldest, least explainable, and least fixable.
Continue reading
Show 4 more in Inventory
Demand forecasting and reorder points that survive reality
Simple reorder rules fail in seasonal, promotion-heavy markets. Here is how to build forecasting logic that accounts for variability rather than ignoring it.
Continue readingPurchase orders need partial receipts from day one
Suppliers short-ship, back-order, and split deliveries. A PO that is either open or received cannot describe what actually arrived.
Continue readingWhat good warehouse picking and barcode workflows look like
Most warehouse software fails at the shelf, not in the dashboard. A look at picking strategies, scan discipline, and the metrics that matter.
Continue readingMulti-channel stock sync and the overselling problem
When one SKU sells on four marketplaces at once, sync latency turns into cancelled orders. Here is where the gaps come from and how to close them.
Continue reading
EdTech
10 notesResults day is a load test you are not allowed to fail
Every student in the cohort loads the same page in the same ten minutes. Average traffic tells you nothing about the only moment that matters.
Continue readingCohorts are not just a group id
A student moves classes in November. Every grade, attendance record, and report already written has to remember which group they were in at the time.
Continue readingOnline exams fail on the network, not the questions
A student's connection drops for forty seconds during a timed assessment. What your system does in those forty seconds decides whether the exam is defensible.
Continue readingGradebooks need a recalculation story
A teacher changes one weighting in March. Every derived grade, predicted outcome, and progress flag downstream is now stale, and nothing tells anyone.
Continue readingBuilding for the student on 2G and a second language
Most of India's next hundred million learners will arrive on constrained networks in a language that is not English, and platforms are still not built for them.
Continue readingSafeguarding changes your permission model
Most systems answer "can this teacher see this student". Safeguarding asks a harder question — who must never see this, and who must always be told.
Continue reading
Show 4 more in EdTech
The engagement metrics that actually predict dropout
Most learning platforms track the wrong signals. The ones that predict a student leaving are usually visible weeks before anyone notices.
Continue readingLMS integrations break in August
Every school system syncs students, staff, and timetables from somewhere else. All of it changes at once, in the two weeks nobody is available to fix it.
Continue readingAssessment integrity when the exam hall is a bedroom
Online proctoring has moved past webcam surveillance toward assessment design, and the shift matters for anyone issuing a credential that carries weight.
Continue readingAI tutoring and the personalised learning path
Adaptive tutoring finally works well enough to matter, but only when the learning path is built on honest data about what a student actually knows.
Continue reading
Next step
Want this kind of thinking on your system?
The posts are the free version. The engagement is the same reasoning applied to your codebase.
Start a project
Replies within one working day