Field-Level Confidence Scoring and the Human-in-the-Loop Review Queue
A well-calibrated per-field confidence score is worth more than a few points of raw extraction accuracy, because it is what lets you route work instead of reviewing everything.
Pricing IDP: Per Page, Per Document, Per Field, or Per Outcome
Four pricing metrics for document processing, what each one incentivises, and why the most defensible choice is rarely the most popular one.
What "97% Accuracy" Means to a Vendor vs. to an AP Clerk
The same accuracy number describes two completely different experiences depending on which side of the software you sit. Bridging that gap is a product responsibility.
Table Extraction Is Still Unsolved: Practical Workarounds That Ship
Line-item extraction remains the hardest part of document processing. Here are the workarounds that survive contact with real customer documents.
Designing the Exception Path: The 3% of Documents That Eat 80% of Support
In document automation the happy path is a commodity. The exception path is the product, and it is where every deployment's real cost concentrates.
ERP Integration Is the Real Product, Not Extraction
Extraction is becoming a commodity. The durable value in document automation sits in everything that happens between a clean extraction and a posted record.
Why OCR Accuracy Is the Wrong Metric — Use Straight-Through Processing Rate Instead
OCR accuracy is a component metric that tells you almost nothing about whether your document automation product is working. Straight-through processing rate does.
The Document Types That Break Every Extraction Pipeline
A field guide to the documents that quietly destroy straight-through processing rates, and what to do about each one.
Template-Based vs. Model-Based Extraction: When the Old Way Still Wins
Template extraction is unfashionable and frequently the correct engineering choice. A framework for deciding which approach a given document population deserves.
Building a Golden Dataset When Every Customer's Documents Differ
How to construct and maintain an evaluation set for document extraction when your customers' document populations have almost nothing in common.
Overview of the Intelligent Document Processing (IDP) Industry
A field overview of the Intelligent Document Processing industry: market, technologies, competitive landscape, workflow, applications, challenges, and outlook.
Model Deprecation as a Product Risk: Writing a Vendor Migration Plan
Your product depends on a component your supplier will retire on their schedule. Treating that as a planned event rather than an emergency is a design decision.
What to Centralize vs. Federate When Building an Internal AI Platform
A dividing line for internal AI platform teams: centralise what is expensive to duplicate and dangerous to get wrong, federate everything that touches the domain.
Fine-Tune, Prompt, or Just Add Rules — A Decision Framework
Three ways to change an AI system's behaviour, with wildly different costs and reversibility. A practical order of preference and the signals that justify moving up it.
Evals Before Features: Regression Testing a Non-Deterministic System
You cannot ship AI features responsibly without an evaluation harness, and building it after the first feature is how teams end up frozen.
Latency Budgets for AI Features Users Will Actually Wait For
How long users tolerate an AI feature depends less on milliseconds than on what they are doing and what they can see. A framework for setting and defending a latency budget.
Hybrid Routing Between Cheap Models, Expensive Models, and Humans
A three-tier routing architecture and the hard part nobody talks about: producing a difficulty signal good enough to route on.
Structured Output and the Failure Modes JSON Schemas Hide
Constrained decoding guarantees your output parses. It guarantees nothing about whether the output is true, and that gap is where production incidents live.
Token Economics at Document Scale: The Cost Model Nobody Plans For
A cost model that is invisible in a prototype and dominant at ten million documents a year, plus the levers that actually move it.
Prompt Versioning as a Release Process
Prompts are production code with none of the safeguards code gets by default. Treating them as releasable artefacts is what makes AI features maintainable.
Migrating a Production Pipeline to LLMs Without Breaking Customers
A staged migration pattern for replacing a deterministic production component with a language model, without discovering the regressions in your customers' month-end.
Designing Handoff: When the Agent Should Fetch a Human
Handoff is not the assistant failing. It is a designed transition with triggers, context transfer, and a return path — and most implementations get all three wrong.
Guardrails That Don't Lobotomize the Assistant
Most guardrail systems are built by accretion after incidents, and end up producing an assistant that refuses everything interesting. There is a better structure.
Context and Memory in Support Agents: What to Persist, What to Drop
Agent memory is usually framed as a capability. It is better understood as a set of retention decisions, each with a cost and a privacy consequence.
Your API Is a Bad Agent Interface — Tool Design for Agents
Exposing your existing REST API as agent tools is the obvious move and usually the wrong one. Tools are a distinct interface with different design rules.
Agent Observability: The Traces You'll Wish You Had Logged
When an agent does something inexplicable in production, your ability to explain it was decided months earlier by what you chose to log.
Letting Your Ticket Taxonomy Drive the Agent Roadmap
The most reliable prioritisation input for an in-product assistant is already sitting in your support system, badly categorised and largely unread.
Deflection Rate Is a Vanity Metric — Measuring Assistant Value Properly
Ticket deflection is easy to measure, easy to inflate, and disconnected from whether your assistant is helping anyone. Better instruments exist.
Lessons From Running a Customer-Facing Assistant at 500+ Conversations a Week
What changes when an in-product assistant stops being a demo and starts carrying real conversation volume, week after week.
Multi-Step Agents vs. Deterministic Workflows
Agentic loops are the default architecture in a lot of AI products and the correct one in a minority of them. A decision test based on how much of the process you already know.
Making a Company AI-Native: Sequencing Internal Adoption Before Customer-Facing Bets
The fastest route to good customer-facing AI runs through internal adoption first — not because it is safer, but because it is how an organisation learns to judge quality.
Certification as a Moat: The Unglamorous Regtech Advantage
Certifications, accreditations, and registrations are slow, expensive, and boring — which is precisely what makes them defensible in a market where software features are not.
Roadmapping When Regulators Set Your Release Dates
Regulatory deadlines are the one commitment a product team genuinely cannot move. Planning around them requires a different structure than ordinary roadmapping.
The Cost of Getting One Validation Rule Wrong
A single incorrect rule in a compliance product does not produce one bug. It produces a fan-out across customers, documents, and time that is unlike anything in ordinary software.
What a Verifactu Rollout Actually Takes
The engineering, product, and operational work behind a billing-software compliance rollout of the Verifactu type — and where the schedule actually goes.
Peppol in Plain English for Product Managers
What Peppol actually is, what the four corners mean, and which parts of it become your product's problem. Written for PMs, not for standards committees.
Audit Trails as a Feature, Not a Logging Concern
Application logs answer an engineer's questions. An audit trail answers an auditor's. They are different artefacts, and treating the first as the second fails at exactly the wrong moment.
Building for E-Invoicing Mandates When Every Country Differs
An architecture and an operating model for supporting many jurisdictions' invoicing rules without a separate codebase per country.
A Rules Engine Non-Engineers Can Safely Update
Compliance rules change on regulatory timelines. If changing one requires a deployment, you are permanently coupled to somebody else's calendar — but naive configurability is worse.
Sold to One Buyer, Used by Another: The Regtech Split
In compliance software the person who signs and the person who uses it daily want opposite things. Serving only one of them is the most common way these products fail.
Compliance Deadlines as Your Go-to-Market Calendar
Regulatory deadlines give you something no other market provides: a published date when a defined set of buyers must act. Building the whole commercial motion around it.
Why Customer Interviews Mislead You on AI Features
Standard discovery technique breaks down for AI features in specific, predictable ways. What to do instead of asking people what they want.
What PMs Should Personally Build to Stay Credible
The argument for product managers building things themselves, what to build, and where the practice turns counterproductive.
Senior PM to Director: What Actually Changes in the Job
The step from senior PM to Director is not more of the same work. It is a change in what you are held accountable for, and most of the difficulty is in letting go of the old job.
Managing an AI Feature's Failure Budget With Sales and Support
Every AI feature has an error rate you have chosen. Making that choice explicit, and shared with the teams who absorb it, is what prevents it becoming a trust crisis.
Making the AI Investment Case to a CFO
A CFO is not evaluating your technology. They are evaluating a cash flow with a risk profile, and most AI business cases are written in the wrong language entirely.
Writing a PRD for Behavior You Can't Fully Specify
A traditional PRD describes what the system will do. For an AI feature you cannot enumerate that, so the document has to specify something else — and here is what.
Roadmapping When Model Capability Is Your Biggest Dependency
Planning around a dependency that improves on someone else's schedule, in directions you cannot predict, without either betting the roadmap on it or ignoring it.
Hybrid Revenue Models Without Confusing Customers
Subscription plus usage plus hardware plus services is where most real businesses end up. Keeping that comprehensible is a product design problem, not a pricing-page problem.
Pricing an AI Feature Inside an Existing SaaS Product
You have an existing price book, existing contracts, and a new feature with real marginal cost. The four options and how to choose between them.
The AI PM's Evaluation Toolkit
The evaluation methods a product manager should personally understand and use — not the research versions, the ones that decide whether a feature ships.
Hardware Plus SaaS: The Hardest Business Model, and the Most Defensible
Why connected-hardware businesses are punishing to operate and extremely difficult to displace once established — and what has to be true for the trade to be worth it.
What a Hardware+SaaS Path to a ₹100 Crore Valuation Requires
Working backwards from a valuation target to the revenue, deployment, and operational realities that have to be true first.
BOM Cost vs. Subscription Price: Connected-Device Unit Economics
The arithmetic that decides whether a connected-device business works, including the costs that never appear in a bill of materials.
Channel-First GTM: Selling Through AMC Contractors, Not End Customers
In industrial markets the contractor who already maintains the equipment is a better route to the customer than your own sales team. Building that channel is a different discipline.
White-Labeling Your Dashboard: What to Expose, What to Hold Back
Partners will ask for the product under their own brand. Where you draw that line determines whether you have a channel or a disintermediation problem.
Connectivity and SIM Strategy for a 4G Product in India
Choosing how a fleet of cellular devices connects in the Indian market — the options, the regulatory realities, and the decisions that are hard to reverse.
Pilot to Deployment: The Valley Where Industrial IoT Startups Die
Successful pilots that never become deployments are the characteristic failure of industrial IoT. The causes are structural, and most of them are addressable before the pilot starts.
Financing Hardware Inventory While Bootstrapped
The cash cycle is what kills bootstrapped hardware companies, not the product. The levers available to shorten it, ranked by how much they actually help.
Warranty, RMA, and the Support Costs Software People Forget
When your product is a physical object, support involves moving objects. The cost structure that follows surprises every team coming from software.
Selling a Device Where the Software Is the Product
Buyers evaluate the object because it is the part they can see. Getting them to value the software instead is a positioning problem with a specific set of moves.
ESP32-S3 for Production, Not Prototypes: A Checklist
The gap between a working ESP32 prototype and a device you can ship a thousand of, expressed as the checks that actually catch problems.
Power Design for Field Devices — LiFePO4, Charging, Brownout Survival
Power is the leading cause of field failure in battery-operated devices, and most of it is decided before any firmware is written.
Breadboard to KiCad: A Product Manager's First PCB
What actually happens when someone who specifies products for a living designs their first circuit board, and what the experience is worth.
Battery Pack Basics: 18650 Configurations and BMS Selection
How series and parallel configurations work, what a battery management system must do, and the decisions that determine whether a pack is safe and lasts.
Debugging Firmware in the Field With No Serial Cable
Once a device is deployed, your only debugging instrument is what it chooses to tell you. Designing that instrument is a development-time decision.
What a 22-Node ESP-NOW Mesh Taught Me About Field Networking
Lessons from building a multi-node local wireless network — where the theory holds, where physical reality wins, and what I would design differently.
What a Secure Element Actually Buys You in Cheap Hardware
A dedicated security chip solves a specific problem well and several adjacent problems not at all. Knowing which is which determines whether it is worth the cost.
OTA Updates for Devices You Can't Physically Reach
Over-the-air update is the capability that determines whether a deployed fleet is maintainable. The design is mostly about what happens when it goes wrong.
SIM7600 vs. A7670C vs. EC200U for Indian Deployments
How to choose between the common LTE modules for an Indian industrial deployment, and which criteria actually decide it.
Deep Sleep Is a Product Decision, Not a Firmware Trick
How often a device wakes determines its battery life, its responsiveness, its data quality, and its cost. That is a product trade-off being made by whoever writes the sleep logic.
Choosing an MQTT Broker: Managed, Self-Hosted, or Serverless
The broker decision looks like a comparison of throughput figures and is really a comparison of operational burden, cost curves, and how much control you need.
The Real Monthly Cost of a "Cloud-Connected" Device
A full accounting of what one device costs you every month after it is installed — including the lines that never appear in the original model.
Alerting Without Alarm Fatigue
An alerting system that produces alerts nobody reads is worse than none, because it creates the appearance of monitoring. Designing for attention rather than coverage.
Provisioning and Pairing Flows a Non-Technical Installer Can Complete
The commissioning flow is where IoT deployments actually fail. Designing it for the person who will be standing on a ladder is a first-order product problem.
Multi-Tenancy for Device Fleets
Device fleets have a harder tenancy problem than ordinary SaaS, because devices move between owners, sites nest, and partners need visibility across customers.
Time-Series Storage for Telemetry in Practice
Telemetry storage decisions made in month one determine your cost and query performance in year three. The choices that matter and the ones that do not.
Telemetry Schemas You Won't Regret in Year Two
The message format is embedded in deployed firmware, which makes it the most permanent decision in an IoT product. Designing it for change is the whole discipline.
Supabase as an IoT Backend: How Far It Scales
An honest assessment of using a Postgres-based application platform as the backend for a connected-device product — what it does well and where it stops.
Offline-First Devices: Buffering, Backfill, and Clock Drift
Connectivity is not a property of a deployed device; it is a condition that comes and goes. Designing for that changes the firmware, the protocol, and the backend.
ThingsBoard vs. Rolling Your Own Dashboard Stack
An IoT platform gets you to a working product in weeks and constrains you afterwards. When that trade is right, and how to tell when you have outgrown it.
Capex-Preferring Buyers and How to Structure Around Them
Industrial buyers often have structural reasons to prefer a capital purchase over a subscription. Fighting that preference is less productive than designing for it.
Mapping the CPCB Compliance Landscape as a Product Opportunity
India's environmental compliance regime creates recurring obligations for tens of thousands of industrial facilities. Reading that as a market map rather than as regulation.
Service Networks and the Support-Radius Problem
A hardware product can only be sold where it can be serviced. That constraint shapes geography, growth rate, and partnership strategy more than demand does.
Selling Proof, Not Sensors: Compliance Evidence as a Dashboard
In compliance markets the customer's real purchase is defensibility. Designing the product around the moment of proof rather than around measurement changes everything downstream.
Reading a Regulation as a PRD Input
A regulation is a requirements document written by people who were not thinking about software. Translating one into a product specification is a distinct skill.
Fire Safety Compliance in Indian Industry: Who Buys, Who Is Liable
Fire safety obligations in Indian industrial facilities involve several parties with different exposures. Understanding who carries the risk determines who will pay.
The FLA and AMC Contractor Channel: Economics and Incentives
Licensed agencies and maintenance contractors are the natural route to industrial facilities in India. Their economics determine whether a channel programme works.
Tendering and Procurement for a Small Hardware Vendor
Public and large-enterprise tenders are written to exclude small vendors, usually without meaning to. Deciding which to pursue and how to compete.
Building for Inspectors: The Second User of Every Compliance Product
Someone who never bought your product, never logged in, and has authority over your customer will eventually judge its output. Designing for them is a real product requirement.
OCEMS and Diesel Genset Monitoring: What the Rule Actually Requires
Continuous emission monitoring obligations create a specific technical and evidentiary requirement. Reading it as a product specification rather than as a compliance summary.
Building Tools for Yourself First
Solving your own problem is the most reliable way to build something good and the most reliable way to build something nobody else wants. Both are true.
Compressing the Solo Founder Workflow With AI
Where AI tools meaningfully change what one person can do alone, where they do not, and the specific failure mode of over-relying on them.
Twelve Years of Shipping: The Mistakes I Keep Repeating
The errors that recur despite being known, why knowing about them is insufficient, and the structural fixes that actually work.
Validating a Hardware Idea Before Paying for a PCB Run
Hardware validation is harder than software validation because iteration costs money and weeks. The cheap tests that resolve most of the uncertainty first.
Idea to PRD in a Weekend: Template and Process
A two-day process for turning a vague idea into a document good enough to build from or kill — and the sections that do the real work.
Running a Product Job and a Hardware Venture in the Same Week
How two demanding pieces of work coexist, what has to be given up, and the specific failure modes of doing both badly.
What Building an ESG SaaS Taught Me About Exit Options
Exit options are determined by decisions made years earlier, most of which do not look like exit decisions at the time.
Day Theming and a Personal Operating System for Builders
A personal system that survives contact with a demanding week — built around modes rather than task lists, and around removing decisions rather than adding structure.
Designing a Solar-Powered Water Dispenser for Animals
A small project with real constraints — no mains power, no maintenance, animals as users — and the design decisions each of those forces.
Killing a Product Idea: The Criteria I Use
Stopping is harder than starting and more valuable. Pre-committed criteria, honest signals, and how to end something without losing what it taught you.