Kgotso Msiza@kgotsomsiza
🤓 Latest Submissions
.png&w=828&q=75)
Aperture: a desk that fires its own strategies
Most "AI trading agents" are one model call wearing a hat: imagine a trade, decide it is safe, place it. Aperture splits that into roles that can disagree, because a desk where nobody can say no is not a desk. Four agents hold real decision rights. Scout picks the day's universe from implied versus realised volatility and the news, but only from a vetted liquid list. Regime sets the posture. Red Team argues against every proposal, seeing the structure but not its author, and needs a concrete objection to kill. Portfolio Manager ranks survivors by conviction, which scales size downward only. One component holds the veto, and it is not a model. Before an order exists, a deterministic Risk Warden reduces every leg to expiry payoff geometry and computes worst-case loss. One uncovered region is an automatic veto. It refuses stale quotes, wide markets, thin depth, concentration, a 3% daily loss and an 8% drawdown. The agents never emit a strike, a price or a quantity. Across the scored window it refused 75 proposals and the Red Team killed 29 more. Both Alpaca surfaces are used for what each is good at. MCP is the eyes, and the only one reaching news; it tags its output untrusted and Aperture honours that, so a headline is data, never a prompt. The CLI is the hands: synchronous, exit-coded, idempotency-keyed, submitting atomic multi-leg orders. Strategies are employees. Realised return per dollar of risk earns budget; repeated vetoes count as miscalibration before money is lost. Two of the four sleeves were fired mid-week for that reason, autonomously. The result we are proudest of is a refusal. A 27-configuration sweep found a variant at +20.0% edge, t=4.89, clearing a multiple-testing bar that rises with every variant tried. On an embargoed holdout it had never touched, it scored t=0.95. It was not promoted. Three research runs, 24 trials, zero promotions. Final scored result: $98,918.89, -1.081%, reported at the official close rather than a later revaluation.
4 Sep 2026
Bastion: Accuracy-First Token Router V23
Bastion is an accuracy-first, token-efficient general-purpose AI agent for Track 1. It classifies every request into one of eight capability categories, then uses a three-tier routing policy. Narrow, provably safe forms are handled by deterministic code at zero Fireworks tokens. Eligible factual, sentiment, entity extraction, and code tasks can run on a bundled Qwen3.5-4B model; Bastion accepts a local answer only when generated-token confidence and a category verifier both agree. Everything uncertain, rejected, or truncated escalates through Fireworks. Gemma 4 31B is prioritized for multi-step reasoning, knowledge, and summarization, while Kimi specializes in code and provides failover. For difficult reasoning tasks, the model works through the problem while Bastion submits only the clean final answer. Completion caps, retries, model failover, truncation recovery, and strict output validation keep the batch resilient. Bastion protects accuracy first, then spends tokens only where they add value.
13 Jul 2026
AutoReview Crew
AutoReview Crew turns code review into a visible multi-agent workflow. A human mentions the Lead Reviewer in a Band room, and the Lead coordinates a full pull-request review through Band: it delegates to Correctness and Security specialists, waits for the current round of reports, recruits a Test Reviewer when coverage is required, aggregates every finding, and escalates risky changes to the human owner. The project targets Track 2: Multi-Agent Software Development. It covers automated PR review, merge preparation, test coverage, cross-model code review, and human-in-the-loop escalation. Band is not a wrapper around the workflow; it is the collaboration layer. @mentions route the work, the shared room carries context between agents, Band participant tools let the Lead recruit a specialist mid-review, and room events provide an audit trail of every tool call and decision. The demo run completes in about 51 seconds. The Security Reviewer, running Qwen2.5-72B on Featherless AI, flags a hardcoded Stripe API key, SQL injection, and missing input validation. The Correctness Reviewer, running through AI/ML API, flags division by zero and a mutable default argument. The Lead then recruits the Test Reviewer through Band, receives concrete pytest coverage, and only then sends its final ESCALATE_TO_HUMAN verdict. The crew is cross-model and cross-provider by design: AI/ML API powers the Lead, Correctness, and Test agents, while Featherless AI powers the open-source Security Reviewer. The result is not one chatbot reviewing code, but a team of specialists that coordinate, recruit help when they need it, and know when to hand a risky change back to a human. MIT licensed.
19 Jun 2026
