Better trades.

A $40B energy conglomerate needed sub-4% MAPE on ERCOT day-ahead load forecasts to run profitable congestion revenue rights. Legacy vendor models sat at 5-7% error — enough to turn winning bids into losses.

We deployed TS2Vec temporal embeddings paired with hybrid CNN+LSTM ensembles. The architecture treats each node as a state-space model where load, weather, and price interact as hidden variables. 200+ models went live within 90 days, each tuned to a specific congestion point.

The same forecasting engine now extends into industrial demand response and datacenter load intelligence, where sub-hour accuracy drives curtailment timing and capacity planning.

3.21% MAPE on day-ahead forecasts. The spread between forecast accuracy and bid profitability compounds daily.

Healthcare revenue cycle management leaks 3-5% of net revenue through denials, underpayments, and manual claim statusing. We structured a joint venture with a national RCM operator to deploy AI across the full cycle — eligibility verification and prior authorization, claim submission, denial prediction, and appeals.

The model maps 112 discrete processes, scores each for automation readiness, and sequences deployment by revenue impact. Denial prediction alone catches 34% of preventable denials before submission. Claim statusing — the largest labor cost in RCM — runs 41% faster with structured extraction.

JV structure means our upside compounds with theirs. 2x EBITDA at the 3-year mark, 3x terminal value on exit.

Equipment lessors price residual values using depreciation tables and analyst judgment. The spread between predicted and realized residual is where margin lives — or dies. A mid-market lessor with $2.8B under management needed better residual estimation across construction, medical, and transportation fleets.

We deployed state-space models as hidden-variable estimators — treating equipment condition, utilization cycles, and secondary market liquidity as latent states inferred from observable transaction data. The model ingests auction results, OEM lifecycle data, and macroeconomic signals to produce residual distributions, not point estimates.

The same state-space architecture powers our energy price forecasting and healthcare process models — hidden-variable estimation applied to different domains.

Underwriting dropped from 9 days to 40 minutes. The model doesn't replace the credit committee — it arms them with distributions instead of guesses.

A $250M listed paper mill runs recovery boilers, lime kilns, and digesters — each a continuous process where energy consumption drifts with feedstock variability, ambient conditions, and equipment degradation. Traditional process control optimizes each unit locally. The savings live in cross-unit coordination.

We deployed graph neural networks for trim optimization and digital twins for recovery boiler operation. The state-space model treats the entire mill as a coupled system — steam header pressure, black liquor solids concentration, and electrical load interact as latent variables. The optimizer recommends setpoint adjustments every 15 minutes.

This is the same industrial power cost optimization framework we deploy across energy-intensive verticals — the physics changes, the architecture doesn't.

8% energy savings verified at scale. In a $250M mill, energy is the second-largest cost line. The math writes itself.

A third-party administrator processing group medical insurance for corporate clients could not act on its own files. Records arrived as PDFs, scans, phone photographs and handwriting, and every enrolment waited on a person to read one. The backlog ran to 25,000 documents going back two years and grew every month. For a business paid a fixed fee per member, that reading was margin leaving the door.

We started with one question: can the file be turned into structured data reliably enough to decide on. We digitised the document set, mapped every extracted field onto a medical taxonomy, and validated the output against the application forms the administrator already held. 353 of 385 reports came through the first validation run. Accuracy across all reports settled at 98%.

Production followed in four stages over seven weeks: infrastructure, model fine-tuning with a human-in-the-loop queue for exceptions, the historical backlog, then an API for real-time submission intake. On top of the extraction sits the piece that moves the number, a rule engine deciding which files clear straight through and which need a human. The pipeline reads 60,000 pages a month.

Cost to serve stopped scaling with the book. The next corporate account does not cost what the last one did.

A consumer-rights claims firm running PCP, Plevin and irresponsible-lending cases could not open new claim categories at the speed its market was moving. Each claim arrived as credit agreements, lender statements and correspondence in whatever format it had been sent. A person read every one before a claim could be prepared, and the firm wanted to be in twenty-plus categories. Reading was the whole constraint.

We started on a specialist document-processing platform, which is what the category tells you to buy. It was the wrong answer. General-purpose models had crossed the line where they read these documents better than the specialist stack and cost a fraction of it, so we rebuilt the extraction layer on Gemini and left the rest of the pipeline standing. Per-page cost landed near a seventh of what conventional OCR pipelines charge.

The pipeline is deliberately plain. The firm stages documents in a shared S3 bucket, extraction pulls the fields that matter for each category, a human-in-the-loop queue takes anything the model flags, and validated records return in a format the firm's CRM ingests without a middle step. Pricing runs on throughput rather than seats, tiered to seven and a half million pages, so the cost of opening a category is a line the firm can forecast before it commits.

Reading stopped being the reason a category stayed closed.

Almost every company in ecommerce is private. Nine million of them, and three hundred and fifty thousand direct-to-consumer brands on top, publish nothing about revenue, order volume or growth. The investors, strategy firms and platforms that need those numbers were buying panels and surveys, which sample a few thousand companies and extrapolate to the rest. The sample is never the companies you are actually asking about.

PipeCandy modelled the population rather than sampling it. We tracked 150 parameters for every company in the universe, from storefront technology and catalogue behaviour to shipping, payments and traffic signals, then ran an ensemble of machine learning and deep learning models across them to predict revenue and order volume for companies that had never reported either.

The test the market cared about was whether a model could call a public company's quarter before the company did. We predicted Shopify's revenue ahead of its quarterly results, accurate to 97%. The same machinery then went down a level, running nightly sales predictions on 45 million SKUs, the work that sits behind product and market intelligence and demand forecasting today. Hedge funds, private equity firms, BCG, Bain, EY and Shopify itself bought the output.

The companies never reported a number. We were right about them anyway.