Machine learning development in New York
New York's insurers and brokers sit on years of claims history: structured fields, adjuster notes, photos and correspondence. The modelling that exists often lives in an actuarial team's spreadsheets or one data scientist's notebook, run by hand when a question comes up. Meanwhile the claims desk triages by reading every file, and the renewals team decides which accounts need attention from memory.
We build the models that turn that history into something the desk uses every day: a triage score for new claims, a flag for the ones likely to escalate, and clustering over adjuster notes to surface repeat causes, the same approach we used on work-order text to find repeat faults for FM360. Training data is versioned so any result can be reproduced, inputs and outputs are logged, and the model is watched for drift as the book changes. Anything that affects a claim decision stays a recommendation to a person, shown with the reasons behind it.
Most software bought in New York is bought by companies that do not think of themselves as software companies. Banks, insurers, asset managers, publishers, agencies, fashion houses and property firms all run on systems that sit beside the real business: a client portal, a pricing engine, a reconciliation job, a content pipeline. Those systems rarely start from nothing. They connect to a vendor platform, a core system older than the team, and a folder of spreadsheets someone updates by hand, and the brief is usually to make all three agree.
What those companies need built follows the industry. Asset managers and insurers need client portals where a family office sees two funds and not a third, and overnight jobs that reconcile custodian, market data and order feeds before the market opens. Publishers and ad-tech firms need event pipelines whose delivery numbers agree across vendors, and front ends that stay fast under the ad stack. Property firms need apps for technicians working in basements with no signal. And founders who left one of these industries need a first version in front of a pilot customer before the interest cools.
Hiring is the constraint. Engineers here are courted by banks, big tech offices and funded start-ups at the same time, and Built In puts the average software engineer base salary in the city at about $160,000 before bonus. The shape this produces is a small in-house team that keeps the core systems running, and behind it a list of well-defined projects that nobody on that team will reach this year. That list is where we are useful: each project is a defined piece of work, it starts with a fixed-price two-week piece, and the code belongs to you from the first commit.
We are in Bengaluru and move our working day for New York. Four hours of every working day overlap with your morning in Eastern time, stand-up included, so decisions and code review happen live with the engineer who writes the code. Runbooks and architecture notes are written as we go, so your in-house team can run what we build long after the project is finished.
We are the right fit for the defined projects your in-house team is too stretched to reach this year: a client portal, a reconciliation job, an event pipeline, a first version for a pilot customer. Senior capacity starts within days, the first piece is a fixed-price two-week engagement, and agencies can ship the work under their own name.
Four ways this arrives.
Someone got good results on a laptop and nobody can reproduce them. We turn it into a training pipeline with pinned data and versioned artefacts, put a serving endpoint in front of it, and record what it was evaluated on so the next version can be compared.
Hundreds of conditions routing tickets or flagging risky orders, each added after an incident. We measure what the rules get wrong on your labelled history, then test whether a classifier beats them. Sometimes the answer is fewer, better rules, and we say so.
Demand, load or staffing estimates that miss often enough to be ignored. We start with a plain seasonal baseline, backtest against the periods you actually plan around, and add complexity only where it beats that baseline on those periods.
Duplicate records, recurring faults, near-identical products under different names. Embeddings group text by meaning where keyword matching cannot, and we build the screen that lets a person correct a group rather than trust it blindly.
Repeat-fault detection that began as UMAP and HDBSCAN over 33,000 work orders and shipped as nightly incremental clustering on the same embedding model.
Read the write-up →Asked by New York teams.
How do you work with teams in New York?+
We are in Bengaluru and move our working day for New York, so four hours of every working day overlap with your morning in Eastern time, stand-up included. Calls, reviews and decisions happen in that window, and the rest of the conversation runs in your tools: Slack, GitHub, Linear. The person on every call is the engineer who writes your code, and the code sits in your repositories from the first commit.
Do you build for New York financial firms?+
Yes. The work is usually the systems beside the trading or underwriting desk: overnight jobs that reconcile custodian, market data and order feeds before the open, client portals where investors, brokers and their accountants each see exactly what they should, and internal tools that replace a pricing or renewals workbook only one person understands. We build them to be idempotent and traceable, so every figure can be followed back to the file it came from.
How does your rate compare to hiring in New York?+
Built In puts the average software engineer base salary in New York City at about $160,000, before bonus and benefits. Our published rate is $35/hour, or $5,400 a month for an embedded engineer, with a $5,000 minimum. There is no recruiting time and no employment overhead, and work starts within days of the call. The first piece is a fixed-price two-week engagement at $2,800, so you judge us on working code.
Can you take over a system another vendor built?+
Yes. In New York that is often a client portal or an internal tool built by an agency that has since moved on, with nobody in-house who knows how it runs. We start by reading the code and running it, then write down how it works and where it is fragile before changing anything. A good first two-week piece makes the riskiest part safe and leaves your team runbooks it can use.
Do we need machine learning, or is a rules engine enough?+
Often a rules engine is enough, and it is cheaper to run, explain and audit. If the people who make the decision today can write it down, write it down. A model earns its place when the rules keep multiplying, the inputs are messy text or images, or the pattern shifts faster than anyone can edit the rules. We check against your history before recommending either.
Why not just call a hosted model API?+
Sometimes you should. With low volume, messy language and no labelled data, a hosted model and a good prompt ship quickly. A trained model wins when per-call pricing starts to hurt at your volume, when latency matters, when data cannot leave your systems, or when you have labelled examples a small model can learn from. We price both paths before you commit.
How much data do we need?+
For classification, usually fewer labelled examples than people fear, and embeddings can group text with no labels at all. Forecasting is the opposite: you need enough history to cover the cycles you plan around, and a single year says little about a yearly pattern. We give you an answer after looking at your data, not before.
What happens when the model gets worse?+
It will, because the world it learned from changes. We log inputs and predictions, compare them with outcomes as they arrive, and alert when accuracy or the input mix drifts past a threshold you agreed. Retraining is a scripted pipeline with an evaluation gate, so a new model replaces the old one only when it does better on the same test set.
Can you build a recommendation system?+
Yes, and we would start without a learned model. Co-occurrence and embedding similarity, served from Postgres or a vector index, beat a most-popular list for most catalogues and are easy to debug. A learned ranker comes after that, once you have enough interaction data to train and evaluate one. A learned ranker is something we build from our work with embeddings and clustering, starting from a simple baseline you can measure it against.