Case studies

How the work holds up.

Real projects, concrete decisions, and honest limits. Each case follows the problem, the approach, and what was built or learned.

01 / Data workflows

Tradings.Guru

AI-assisted market research with clearer data boundaries.

Problem

Stock research meant moving between market data, social discussion, and other sources. Collecting that information manually was repetitive and hard to keep current.

Result

A deployed research platform that brings fragmented information into one interface, with more explicit sourcing and methodology in its options-flow tool.

Limits & lessons

Coverage and freshness depend on upstream providers. Some tools remain heuristic or estimated. This is research support requiring human judgment, not evidence of predictive accuracy or investment performance.

Approach

Brought market and social-data tools into one research workflow using APIs, MCP integrations, and stored context for agents.
Built a shared caching layer to manage repeated upstream requests; the broader workflow supports automated updates and a customer-facing chatbot.
Replaced synthetic options-flow metrics with actual options-chain volume and open interest, removed unsupported sweep detection, and included the methodology and update time in the response.

Evaluation & judgment

Checked where outputs came from rather than relying on feature labels. The source catalog distinguishes provider data, heuristics, and estimates; the indicator-based signal route is separate from the broader AI-assisted workflow.

Explore the project

02 / Data quality & judgment

StreetScope AI

Testing a local-market opportunity workflow before scaling it.

Problem

An opportunity map can mistake missing business records for an underserved market. Useful recommendations need coverage and category quality behind the score.

Result

A regional prototype with substantial Florida coverage and a reusable public scraper. The product is paused after nationwide collection and ongoing data costs proved difficult to justify.

Limits & lessons

The scores are heuristics, not validated predictions. Validate coverage, cost, and the usefulness of a score before scaling an automation; a missing listing is not proof of an underserved market.

Approach

Built business collection, normalized categories, database storage, and a map interface for exploring local-market gaps.
Used distance from competitors, nearby commercial activity, and competitor ratings as inputs to explainable opportunity scores.
Simplified the approach when the initial scoring was not useful enough, focusing on more relevant inputs rather than adding complexity.

Evaluation & judgment

Compared the usefulness of the output with the completeness of the inputs. Result caps, uneven geographic coverage, and shallow categories made missing listings an unreliable basis for confident site recommendations.

Explore the project

03 / AI-assisted build & test

Cash Carnage 2

A browser game built through a reproducible bug, test, fix, and playtest loop.

Problem

Real-time co-op has to stay understandable under pressure: players need clear attack warnings, consistent controls, and a shared game state.

Result

A public, no-install co-op arena shooter with a playable campaign and a repeatable test-and-refine development loop.

Limits & lessons

Local multiplayer checks and playtesting do not establish exhaustive device, browser, or real-world network coverage.

Approach

Built room-code co-op, a six-round campaign, an eight-boss roster, and optional Overtime, with desktop and touch controls.
Matched boss charge-warning geometry to the swept collision path so the visible warning describes the actual danger.
Fixed duplicate rush-end projectile volleys and added regression coverage for attack cues and boss behavior.
Used automated regression tests and local four-player checks alongside hands-on playtesting before publishing updates.

Evaluation & judgment

The shipped build passed TypeScript checks, its production build, automated regression tests, and local four-player smoke checks. Hands-on playtesting checked ordinary gameplay after the fixes.

Explore the project

Browse all projects