Projects

Designed and built personally, on personal time and infrastructure. Internal tooling built for an employer stays with that employer and is not described here.

Four packaged research agents

Company deep dive, cited research, entity resolution and counterparty mapping. Forty-nine documents to date. Read onClose

Agent	What it produces
Company deep dive	A complete company file with every claim tied to a named source and a quoted line. Returns in an afternoon what desk research takes days to assemble.
Cited research	A written answer to a specific question, sourced well enough to drop straight into a memo without re-verifying it. The source trail is the deliverable, not an appendix to it.
Entity resolution	Resolves a brand or trade name to the legal entity that actually holds the contracts. That is the step that makes registry, filing and litigation searches return the right company.
Counterparty mapping	Establishes who a company actually does business with when its own materials will not say. Confirms or breaks a claimed commercial relationship before it gets underwritten.

Forty-nine research documents to date.
The largest is a market-structure study running to roughly thirty pages with 76 footnotes.
A full 21-company comparable universe was built and normalized in one sitting rather than accumulated over weeks, so the peer set is internally consistent.

The point of the agents is that the expensive part of research, finding the source and proving the claim, stops being the bottleneck.

A design system

Board decks, investor materials, CIMs and teasers in one house style, with nobody formatting a slide. Read onClose

The division of labor. The writer supplies the content and the argument; the system supplies everything else.
The result. A document that looks like it came from a firm rather than from a laptop, produced in the time it takes to write it.

A model-vetting harness

Puts a model through a structured adversarial review before it reaches a decision-maker. Read onClose

Five checks, each against a fixed rubric. Assumptions, fundamentals, formula integrity, presentation, and scenario and sensitivity coverage.
Findings by severity, location, evidence and recommendation, re-checked round over round so a closed finding cannot quietly reopen.
Catches what reading never does: the plausible number produced by a broken formula.

The purpose is narrow. A model gets reviewed by something with no incentive to conclude it is fine.

Independent Projects: Agentic Work

Agents produce output faster than anyone can verify it, so the verification got built. Designed and built personally, March – July 2026.

Lines of code delivered	493,214
Of which automated tests	134,781 (27%)
Specs, plans, validation docs	793,201
Commits, 20 Mar – 24 Jul 2026	7,881
Active days of 127	104 (82%)
Longest streak	26 days
Heaviest single day	1,638 commits
Release tags across six repositories	25

Basis: git history across six repositories, 20 March – 24 July 2026. No vendored or minified files; largest single sources are ~2,000 lines of hand-structured code.

The Rig shipped v2.1, audit passed 31/31

866 commits · 91,786 lines · 12-package TypeScript monorepo · 206 test files, suite 845/845 · May – Jul 2026 Lets an AI operate real phones, desktops and browsers with screenshot, video and vision-judged proof of every step. Read onClose

Built around one truth: an automated tester that grades its own homework always passes. So:

It re-reads evidence files off disk before it may report “pass.”
A missing screenshot is a hard failure, not a warning.
Every run is fingerprinted.

Its own audit then caught it passing itself on two unwired checks, so a phase was inserted to close the hole.

ArGaze live in production

1,762 commits · 142,090 lines · 273 test files · v1.0–v2.1 shipped, v2.2 in flight · Apr 2026 – present Prove it happened. On camera. In real time. Read onClose

ArGaze turns the camera a team already carries into an AI-assisted workflow verification step: it checks each physical task against the spec as it happens and files a tamper-resistant, timestamped record retrievable by ID. Three functions:

Capture a physical task on any camera.
Verify it against specification with a vision model that flags mismatches as they occur.
File the result as evidence linked to an order, claim or work-order ID.

Errors get caught while they are happening rather than after. The hard problem was making a phone camera, a browser and a remote auditor agree in real time, and keep agreeing when a paired session’s credentials expire mid-recording.

Ranch AI shipped, nine audited milestones

3,671 commits · 162,582 lines · v1.0 → v5.0 · Apr – Jul 2026 An engineering-enforcement layer that sits between an AI coding agent and anything it can change. Read onClose

Mandatory security gates, live user-flow testing against the running application, and a human sign-off step the agent cannot bypass.

The point is that agent output is fast and confident regardless of whether it is correct, so the gates have to be structural rather than advisory. Runs as the development backbone for every other project here.

The Librarian shipped v4.0, running nightly since 12 June

654 commits · 23,431 lines · Jun 2026 Always-on, fully local code-knowledge index queryable by any agent over MCP. Read onClose

Hybrid keyword and vector retrieval over AST-aware chunks, plus cross-language impact analysis. Runs entirely on the machine: zero data egress.

Kayvo build complete, pre-launch

800 commits · 47,023 lines · 29/30 phases, 77/77 plans · 3 release tags · 28 test suites · Mar – Apr 2026 Two-sided paid-messaging marketplace. Read onClose

Alias-based routing, connected-account payment splitting, read receipts, admin and payout console.

Four of five are one thesis attacked from different directions. One refuses to accept its own unverified passes. One enforces process. One gives agents memory. And ArGaze applies the same logic to physical goods: don’t take anyone’s word for what was in the box.

← Home

Continue

Background Transactions Process ParcelPing Work Examples Perspectives

Contact

Solomon Dworsky · Miami, Florida

Email LinkedIn Resume, PDF

The pages here stop at what was built and what it produced. If you want the detail underneath any of it, ask.

Private equity transaction work, healthcare technology, and applied AI in the investment process.