Assera

Platform and research overview

Assera is the financial-technology platform I am building with AI assistance. It brings together market-data workflows, trading signals, AI-assisted decision support, research tools, and user-facing interfaces. Its development combines software engineering with empirical questions about whether a proposed signal or automated decision is useful and reliable.

I guide the platform’s development, define questions to investigate, and use AI-assisted development and evaluation to test ideas against historical data and observed system behavior. The research covers trading signals, AI-generated decisions, execution assumptions, and the reliability and risk controls needed around them. Results inform what to keep testing, revise, or withdraw.

Research across the platform

The research asks whether market signals remain useful under realistic data and timing constraints, whether AI-directed decisions improve on simple alternatives, and how costs, missing observations, and repeated testing change the conclusion. Historical comparisons, replay evaluations, and retained negative results help guide development.

Reliability and risk controls are part of the same platform: checking whether data is fresh, separating research or paper modes from live authority, and investigating how failures affect users. The work includes software and interface development alongside the evaluation of the decisions those systems make.

The detailed studies below provide selected evidence for this broader work. They include an AI-directed exit comparison, market-warning research, and native-client reliability. Each retains its methods, findings, and limitations.

Supporting study: AI-directed exits

The expanded September study starts with 275 entries, identifies 135 from the AI-era window, and retains 108 quote-eligible entries across 31 dates. Twenty-seven entries in that window are excluded because the required quote coverage is unavailable. Four AI policies produce 432 replay paths; these are repeated policy evaluations on shared observations, not 432 independent market samples.

The primary analysis excludes four pilot dates, leaving 91 entries across 27 dates and eight model-arm comparisons. This separation matters because pilot work can influence later choices. Date-clustered uncertainty treats observations from the same trading day as related.

The result: no promotion

All primary clustered confidence intervals span both benefit and harm. The study does not support promoting an AI policy based on these results. That is a substantive finding: the available evidence was insufficient to justify a stronger claim.

Mechanical alternatives were examined as well. Trailing variants performed worse in the reviewed comparison, while tighter-stop results were uneven. Neither a negative result for one control nor an attractive isolated outcome establishes a generally superior policy.

Timing and execution assumptions

The execution audit records 2,326 new model responses and 662 exact-input reuses. Reuse is restricted to identical input states; once policy paths diverge, subsequent decisions need new inputs. The audit also records 292 responses exceeding the assumed 12-second delay.

This exposes an important gap between a decision replay and a system that can execute in the market. A model can make a retrospective choice using an assumed timing rule while real latency changes the available price. These results are fixed-entry cashflow comparisons, not an executable portfolio track record.

Data and implementation corrections

• A hash mismatch was treated as a reason to abort rather than silently mix artifacts.
• Timestamp-unit and coverage-helper issues required correction.
• Timeouts were not counted as completed model responses.
• Static fallback behavior was distinguished from an actual AI decision branch.
• Missing coverage was retained as a limitation rather than assumed to be randomly distributed.

A separate learning experiment

A related training experiment used 147 supported paths from the starting inventory to construct 124 labeled checkpoints with eight features. Four chronological blocks produced 80 forward predictions across 25 dates.

The fitted models did not outperform these baselines on the reported measures. Confidence intervals remained ambiguous, and numerical warnings were preserved. A final fit on all available data was not counted as additional validation, and the experiment did not establish a deployed policy.

My role in the investigation

I directed the product and research questions and asked for comparisons using actual model prompts and historical decision states. This was AI-assisted work, with historical artifacts documenting the evaluation methods, corrections, and results.

The investigation uses inspectable comparisons, retains corrections, and accepts a result that does not support deployment. The resulting report is empirical evidence with explicit limits, rather than a claim of profitable predictive ability.

What would strengthen the next experiment

A stronger next evaluation would pre-register the comparison, freeze decision and execution assumptions, acquire adequate complete coverage, separate model-development data from untouched evaluation data, and test the whole executable policy. Those are future requirements. They are not retroactively satisfied by expanding a historical replay.