Krym StudioMariana Krym
Case study · Private investment

From four hours of daily research to under two minutes.

A private investment methodology now runs through an AI decision-support system that applies the operator's rules across hundreds of securities, under the same source standards and validation requirements.

4 hours → <2 minutes Daily research and screening time
~300 securities Evaluated under the same methodology
0 partial rankings Incomplete inputs produce no report
The problem

The method worked. Applying it manually did not scale.

The operator had more than 15 years of market experience and a defined methodology. The constraint was the daily work required to apply it with full discipline.

Research meant crossing several trusted data sources, checking each candidate against a sequence of criteria, monitoring existing positions, and keeping track of conditions that changed what deserved attention. The work took up to four hours a day and limited how much of the market could be reviewed consistently.

The goal was not to ask AI what to buy. It was to make the operator's own method explicit, apply it repeatedly, and surface what they needed while keeping every investment decision in their hands.

Before

Manual multi-source research, a narrow review window, and hours spent checking whether each security met the defined criteria.

Required standard

The same rules, source requirements, limits, and exceptions applied across the full target universe without silent substitutions or omissions.

What I built

A decision-support system built around one operating method.

I extracted the methodology from the operator's existing work and structured it into rules, source priorities, validation gates, exception handling, and standardized outputs.

The build coordinates three runs. Each has its own rules and output, but all share one operating standard, and the operator's existing market-data tools remain the source of facts. The system applies that logic to the inputs and prepares the result for review.

System output

One operating screen, plus two separate reads that run alongside it.

The main screen is the daily instrument: a 40-gate matrix run across the operator's scoped universe, positioning every survivor against its pivot. Two smaller runs sit beside it, each answering a question outside the main method. The system prepares all of it. The operator decides.

Daily investment brief. A 40-gate buy-moment candidate scan across roughly 300 securities, with a pivot-range map and a ranked top-10 display; a separate MA reclaim radar across roughly 500 securities; and portfolio monitoring across 17 holdings routed by risk. The system prepares the priorities and flags; the operator decides.

Repurposed from a live deployed run. Client identity, holdings, tickers, and gate configuration are redacted, and counts are illustrative. The structure and controls are shown as they run. The early-signals and portfolio reads are separate runs on their own scoped universes, not the main matrix repeated.

The operating standard

The intelligence sits in the rules the model must follow.

The system is valuable because it follows the discipline already governing the operator's work. AI is allowed to reason over verified inputs. It is not allowed to supply facts, fill gaps, change the methodology, or make the final decision.

Source hierarchy

Defines which sources can provide each input, which source takes priority, and which are not accepted.

Validation gates

Required data must be current, complete, and verified before any classification begins.

Human authority

The operator reviews the evidence and stays responsible for every investment decision.

When a value can’t be verified, the system hands over a decision.

Refusing to guess or silently drop an input is table stakes. What matters here is what happens next: the run pauses and offers a choice — reschedule for when the data settles, which the system books itself, or approve a fallback source outside the standard list. The operator decides; the run continues or defers on that call.

Calibration

Reliability came from testing the failures.

The system was calibrated on real runs, not approved after one successful demonstration.

Early runs exposed the shortcuts a general-purpose AI tends to take: using a provisional value, producing a partial ranking, dropping an unresolved security, or reaching for an alternative source without approval. Each rejected workaround became part of the operating rules.

The logic was then tested against the same source data in separate model and execution environments. Corrections continued until identical inputs produced the same classifications and outputs consistently.

Outcome

More coverage. Less research labor. The same decision authority.

The operator now starts with structured research already prepared under the methodology. Time once spent rebuilding the analysis goes to reviewing what surfaced, examining exceptions, and making the decisions that still require human judgment.

Under two minutes

for a daily process that previously took up to four hours.

Hundreds

securities evaluated under the same defined standard.

One methodology

applied across research, alerts, and portfolio monitoring.

What this is not

This is not an AI stock picker or trading bot.

The system does not predict performance, place trades, or replace the operator's judgment.

It structures the research layer around a proprietary method. It applies known rules to approved data, surfaces what deserves review, explains why it surfaced, and stops when the evidence does not meet the required standard.

The system prepares the decision. It does not own it.

Common questions

How this kind of implementation works.

Does the system make investment decisions?

No. It applies the operator's rules to verified inputs, surfaces candidates and flags, and prepares standardized research. The operator retains every investment decision.

Does it replace existing market-data tools?

No. Those tools remain the source of market data and technical context. The system coordinates the rules across those inputs and produces the custom research output the process requires.

What happens when required data is missing?

Rather than guess or drop the input, the system pauses and hands over a decision: reschedule the run for when the data settles, or approve a fallback source outside the standard list. The run then continues or defers on that call.

Can the same method be used outside investment work?

Yes. The specific investment rules remain proprietary, but the implementation method transfers to other recurring work built on trusted sources, explicit rules, exceptions, validation standards, and human decision boundaries.

Implementation

How an implementation runs.

1 · Map

Document the workflow, sources, rules, limits, exceptions, and review requirements.

2 · Build and test

Structure the system, expose the failure paths, and calibrate outputs against real runs.

3 · Handoff

The operator receives and runs the system, with the review boundary intact.

The investment rules on this engagement are proprietary; the implementation method is not. Scope is agreed first, then a fixed project price before work begins.

Built by Krym Studio

Mariana Krym has 20+ years building and scaling technology platforms, including a consumer platform scaled to approximately 15 million users and a privacy-first personal-AI product. Krym Studio structures proprietary knowledge, rules, and expert judgment into AI systems that produce from a business's own context.

What expert work is still handled manually?

If a valuable workflow depends on proprietary knowledge, strict rules, or expert judgment, I can assess what should be structured and build the system around it.

Let's talk