Kayvon Salari

An equity research system built on one idea: AI you can audit. A company is compared with peers drawn from about 25,000 listed companies and run through a set of stock strategies, each asking one plain question: cheap and good, growing at a fair price, steady income, durable business. A separate check asks whether the profits are real. Verdicts are fully deterministic: the maths judges, the AI writes, and a fact checker reads every sentence against the numbers. Every strategy carries a ten year backtest record per industry, and a blind three model judge panel gives a second opinion with no vote. Public repo, more than 5,000 tests.

Mortgage paperwork goes in; six figures come out, each tied to the document it came from, with anything that does not add up flagged. The decision never touches the model: a Unity Catalog SQL function applies the lending rules from a policy table, and anything short of a clean pass is referred to a person with a drafted case note. I then attacked it four ways. Three attacks bounced off; a note hidden in a genuine payslip fooled the AI on the first attempt. The README is an honest field report on why, and on the code checks that would stop it. Served as an MLflow endpoint and traced end to end.

Analysts burn hours cross-referencing market data against strategy documents. DualLens ranks companies on both at once — three years of stock performance joined with insights RAG-extracted from strategy PDFs — surfacing which names are financially strong and positioned for AI adoption. One query replaces the manual cross-referencing loop.

Per-company investment research takes an analyst 4–6 hours; this LangGraph agent does the gathering in minutes. It pulls real-time market data, news sentiment, and private analyst reports, then produces a Buy/Hold/Sell recommendation with every source cited — auditable enough to check, fast enough to run across a watchlist. The predecessor to Aristos Council, where its single-agent verdict grew into a gated, deterministic decision core.

Mortgage underwriting takes 3–5 days largely because four assessments — credit, income, assets, collateral — queue behind one another. This system runs them as specialist agents under a Supervisor, cutting the cycle to hours while keeping the parts regulators care about: Fair-Lending compliance, PII protection, bias detection, and mandatory human-in-the-loop escalation on every contested file.

Early multiple-sclerosis signals sit scattered across EHR records where no one has time to connect them. This multi-agent system scans records and flags patients showing early risk patterns for neurologist review — with adjustable autonomy, a transparent rationale for every flag, and PHI governance throughout. It surfaces candidates for clinical judgment; it does not replace it.

Keeping up with AGI-relevant research on arXiv is a weeks-long manual effort with partial coverage. A Planner→Discovery→Evaluation agent pipeline automates it, scoring every paper against a standardized 10-parameter AGI framework — literature review in hours, with 3–5× the coverage of manual survey.

A content pipeline with editorial standards built in: a Researcher gathers material, a Writer drafts, a Critic scores against a quality rubric, and a Supervisor loops them until the draft clears the bar. The interesting part is the loop — generation that doesn't ship until an adversarial agent approves it.

Clinicians wait on DBAs for every data question. This Natural Language-to-SQL system lets them ask directly — and makes it safe by classifying every query before execution: READ runs automatically, WRITE requires human approval, UNSAFE is rejected outright. Full audit trails keep it inside HIPAA and GDPR.

City caseworkers lose time hunting answers across fragmented internal sources of unknown currency. MucAtlas (Munich Innovation Challenge 2026 entry) is a five-agent RAG system that connects those sources, validates whether legal documents are still current, detects contradictions between sources, and returns cited answers in seconds — with full data sovereignty via confidential computing. Architecture and design; competition entry.

Programming warehouse robot routes by hand doesn't scale past the first layout change. This reinforcement-learning agent (PPO) learns goal-directed navigation from scratch and generalizes to layouts it has never seen — pathfinding as a learned skill rather than a maintained ruleset.

Betting markets price the top of the game efficiently and ignore the bottom entirely. Tennis Oracle rates every ATP and WTA player from a million matches — tour, Challenger and ITF — and predicts at 65% where the market sits at 68%, but at 70% on women's ITF events no bookmaker prices at all. Verdicts come from surface-adjusted Glicko-2; the LLM only narrates. Speaks the A2A agent protocol, and a blind test of 197 matches showed the model analyst losing to the ratings by 7.6 points, which is why it narrates.