How a finance-tuned smaller model turned overnight investment research into a near-real-time product without giving up reported task quality.
The production constraint
Boosted.ai built an AI portfolio assistant that analyzes more than 150,000 sources across more than 60,000 stocks. The company wanted to process five to ten times more information and move from overnight analysis toward real-time updates.
Its general-purpose language model reportedly cost nearly $1 million per year. High-end A100 and H100 availability also constrained the design.
Boosted.ai, AWS and Invisible Technologies selected the smallest model expected to meet the finance tasks, tuned it and evaluated it across economics, mathematics and related requirements. According to the AWS case, the result was:
Reported Production Results
90% lower
Model cost
Reported in the cited source.
5-10x more
Desired analysis volume
Reported in the cited source.
About 5-10 minutes
Data update delay
Reported in the cited source.
| Metric | Reported outcome |
|---|---|
| Model cost | 90% lower |
| Financial-analysis quality | Similar to the prior general model |
| Desired analysis volume | 5-10x more |
| Data update delay | About 5-10 minutes |
| Hardware requirement | Smaller, more available accelerators |
At unchanged volume, a 90 percent reduction on nearly $1 million would imply a low-six-figure model run rate. Boosted.ai increased processing frequency, so the public percentage does not reveal the actual post-change annual bill.
The optimization target was task IQ, not general IQ
CEO Joshua Pantony described the requirement as finding the “smallest possible model with the highest possible IQ for our tasks.”
That final phrase changes the procurement process. A general benchmark rewards capabilities that an investment-research pipeline may rarely use. Boosted.ai needed accurate extraction, classification, synthesis and finance reasoning across a known source universe.
Task-specific evaluation can compare:
- entity and ticker resolution;
- numerical and temporal accuracy;
- distinction between reported fact and model inference;
- material-event recall;
- source citation and freshness;
- portfolio-relevance ranking;
- unsupported financial claim rate;
- cost per accepted insight.
The smallest model that clears those gates can outperform a broader model economically even if it ranks lower on general tests.
Lower cost changed product frequency
The benefit was not only a smaller invoice. Overnight analysis limits the value of a product in fast-moving markets. Lower model and hardware requirements allowed Boosted.ai to move toward minute-level processing and a five-to-ten-minute information delay.
This is operating leverage:
same or lower spend -> more sources x more frequent analysis -> more timely user value
An enterprise should therefore calculate both avoided cost and newly affordable capability. The bespoke AI business case treats expanded volume, latency and product coverage as part of the return, not as side effects.
Private deployment was part of the design
The tuned system could run inside a customer's AWS VPC. For financial institutions, model economics cannot be separated from data access, retention, residency and audit controls. A nominally cheaper public endpoint may be ineligible.
A smaller model can make isolated or customer-controlled deployment practical on more available hardware. It can also reduce reliance on scarce top-tier GPUs. That does not make the system operationally free. Monitoring, upgrades, security and capacity remain part of fully loaded cost.
Evidence limits
The AWS article is a vendor-supported startup case. It does not name every model, publish raw evaluation scores, disclose the exact post-change bill or separate model savings from serving improvements. “Similar quality” refers to the tested financial-analysis workload, not general capability.
The case is still one of the clearest public examples of a nearly seven-figure general-model line item being reduced through specialization. It sits in the Frontier Model Downshift Index beside Checkr's fine-tuned classifier, CarbonTrail's financial-data pipeline and StudyFetch's 10x speech case.
For governance and measurement, see the private AI bank architecture and AI FinOps framework.
Building a Production-Ready System
Conscious Engines builds bespoke models for high-volume enterprise tasks where private data and a clear quality function can make a smaller model more valuable than a generic one.
We establish the frontier baseline, tune the candidate on domain examples, test financial and compliance failure modes, deploy it within the required boundary, and retain a stronger fallback for novel work. The result is measured in accepted insights and business latency, not benchmark prestige.