AI and machine learning development
Tankar builds machine learning into working software, from the data pipeline to the model to the interface that uses it. The work starts by checking that the problem needs a model at all, then measures the model against a held-out set of your real data before it goes anywhere near production.
A model is a component, not a project
The prediction only matters when a person or a system acts on it. Tankar builds the whole path: the pipeline that prepares the data, the model, the API that serves it, the interface or workflow that uses it, and the monitoring that says whether it still works next month.
Measured before it ships
Every engagement agrees a metric and a held-out set of real data in discovery. Weekly demos show the number, not a slide about it, and the pre-launch write-up records accuracy, cost per prediction, latency and failure modes.
Typical engagements
What this service usually produces, as deliverables rather than adjectives.
- Demand or sales forecasting fed by your transaction history, with a dashboard and alerts
- Classification of documents, tickets or transactions, with human review for low-confidence cases
- Matching and recommendation, such as tenders to suppliers or products to buyers
- Anomaly detection for fraud, quality or operations data
- Computer vision for inspection, counting or document capture
- Model monitoring, retraining pipelines and evaluation harnesses for an existing model
How it runs
The stages this service goes through, and what you see at each one.
Discovery
The decision the model will inform, the data that exists, a baseline without a model, and the metric that defines success. Written estimate within 48 hours.
The written estimate follows within 48 hours of the scoped call.
Design
Data audit, feature design, and an evaluation plan with a held-out set of real data agreed before any modelling starts.
Build
Pipelines, baseline model, iterations measured against the held-out set, and the interface or API that uses the prediction. Weekly demo of the numbers.
Test
Accuracy, cost, latency and failure modes measured and written up. Bias and edge cases reviewed with the people who own the process.
Launch
Model served behind an API with logging, monitoring of drift and a fallback path when confidence is low.
Run
Retraining on a schedule, drift alerts and a monthly report on accuracy and cost in production.
- Team shape
- A pod of a project manager, a machine learning engineer, a backend engineer and a QA engineer, with a designer when there is an interface.
Use cases by industry
Where this kind of work has paid for itself, by sector. Each is a task with a measurable output, not a demo.
Government and public sector
- Matching tenders to suppliers by category, location and value
- Classifying citizen applications and routing them to the right department
- Forecasting demand for public services from historical records
Retail and e-commerce
- Demand forecasting for inventory and purchasing
- Product recommendations and search ranking
- Fraud and return-abuse detection
Logistics
- Estimated delivery times from historical routes and traffic
- Assignment of jobs to partners by skill, area and availability
- Anomaly detection on fleet and fuel data
Fintech and BFSI
- Transaction classification and anomaly detection
- Document extraction for onboarding and KYC
- Credit and collections scoring with explainable features
Healthcare
- Appointment no-show prediction and reminder scheduling
- Triage of enquiries and document capture from forms
How we evaluate
- Accuracy, precision and recall on a held-out set of real data, compared with the baseline the business uses today
- Cost per prediction in production, including data pipeline and serving
- Latency at the 95th percentile under expected load
- Failure modes, bias checks on protected attributes where relevant, and the fallback when confidence is low
What we do not recommend
We do not recommend a model where a rule or a report answers the question, where there are fewer than a few thousand labelled examples for a supervised task, or where the decision must be fully explainable and a simpler scoring method would pass audit. We also do not recommend starting with a custom model when a proven API or a classical model has not been tried.
Stack for this service
The technologies this work is usually built on, and why each one is used here.
| Technology | Why we use it here |
|---|---|
| Python ML stack (scikit-learn, PyTorch) | Classical models first because they are cheap to run and easy to explain; deep learning where the data and the problem justify it. |
| Data pipelines (Airflow, dbt) | Training data and features are produced by scheduled, tested pipelines rather than one-off notebooks. |
| PostgreSQL | Features, predictions and feedback are stored where the rest of the product can use them. |
| Evaluation harnesses | Every model version is scored on the same held-out set, so improvements are measured rather than felt. |
| Docker | Training and serving run in the same image, on your cloud account. |
Selected work
Case studies where this service carried the engagement. Only outcomes with a source are shown.
Own productGovernment and public sectorTenderBazaar: tender discovery and alerts for Indian SMEs
A tender discovery platform that collects listings from central, state and municipal portals into one searchable feed with daily alerts.Services: SaaS product development, Mobile app development and 1 moreRead the case study
Own productTravel and hospitalityTripBNG: a B2B travel booking platform for agents
A booking platform that gives travel agents supplier inventory, quotes, credit limits and settlement reports in one account.Services: SaaS product development, API development and integrations and 1 moreRead the case study
Engagement models that fit
A discovery and proof of value is usually fixed price; production models run on time and materials or a dedicated team.
- Fixed priceA defined scope with a clear end state: an MVP, a website, a well-specified module or an integration.
- Time and materialsEvolving products, research-heavy work such as AI features, and engagements where the backlog is set sprint by sprint.
- Dedicated teamOngoing product development where you want engineers who know the codebase and stay on it, working as an extension of your own team.
Questions buyers ask
What buyers ask most about this service, answered before the first call.
Tell us what you are building.
NDA on request. Written estimate within 48 hours of a scoped call. Reply within one business day.