Kelly Lyons

Partner Quant Program Overview The Partner Quant Program is VisualHFT's equity-based contributor initiative, designed for highly committed developers, quants, and technical builders to earn long-term ownership (equity) in the #VisualHFT platform and startup. It targets individuals seen as potential long-term collaborators or future employees, focusing on co-building the commercial-grade version of the tool. Key Details - Purpose: To assemble a small team of contributor-partners who own modules, contribute significantly to the core product, and help shape VisualHFT into a leading market microstructure analytics platform. - Compensation: Equity-only (no salary, swag, or other compensation at this stage). Vesting starts after 3 months of consistent contributions. Participants sign a Contributor Equity Agreement and IP Assignment. - Eligibility and Requirements: - Open to anyone with passion, discipline, and proven ability to ship high-quality work (e.g., new graduates reverse-engineering the repo or experienced quant devs). - Emphasis on "energy" over résumés. - Contributions must be original, production-grade, and aligned with the project's mission. - Regular contributions (e.g., PRs) are a starting point, but acceptance into the program is selective for those demonstrating deep commitment. - Suggested Contribution Areas: - Quant signal modules (e.g., VPIN, LOB Imbalance, OTR/TOB ratios, latency surfaces). - Infrastructure plugins (e.g., Kafka, FIX, ITCH, PCAP connectors for live or historical streams). - Visual tools (e.g., execution slippage heatmaps, spoofing alerts, quote stuffing tracebacks). - Market insight tools (e.g., auto-tagging engines, anomaly clustering, post-trade diagnostics). - Any original idea fitting the VisualHFT mission (pitch it directly). - How to Get Involved: - Start by forking the repo, submitting PRs, or building prototypes—strong signals get noticed. - Formal application: Details and application link are in the dedicated file on GitHub. - Repository reference: https://lnkd.in/eGxM_wma This program distinguishes casual open-source contributions (welcome from anyone) from equity-earning partnerships reserved for accepted Partner Quants. It's aimed at those with "skin in the game" who want full access to building the advanced commercial edition alongside the open-source core. For the latest or to apply, check the GitHub repo directly, as the program evolves with the project.

11

A surprising performance dynamic has emerged: CPUs potentially outperforming FPGAs in High-Frequency Trading (HFT) strategies. DeepSeek 3.2's findings suggest a re-evaluation of traditional approaches. Their recommendation? Optimizing HFT code using C++ and meta-templates for maximum efficiency. This shift underscores the importance of continuous testing and adaptation in the fast-paced world of algorithmic trading. #HFT #CPUs #FPGAs #DeepSeek #AlgorithmicTrading #C++

1

DSV Fund continues its investment in the decentralized AI ecosystem. Our investment in Synth, a company that uses AI for predictable intelligence. We believe Synth will become a crucial source of synthetic data for price forecasting, particularly valuable for financial markets and portfolio management. For more information about DSV or investing in open-source AI, please reach out to me. #DSVFund #Cryptoinvesting #DeFi

6

General Tensor has closed $5M across pre-seed and seed rounds. Seed anchored by Good Morning Holdings (a fund backed by Goldman Sachs). Pre-seed led by Lvna Capital, with DCG, X Ventures, Proof of Talk & Outliers Fund. Decentralized AI infrastructure is getting built. https://lnkd.in/eijAyzWP

13

1 Comment

Remember the $2B AUM hedge fund I wrote about? the $500K "Kill Switch" budget, the fully autonomous Design → Code → Test → Deploy loop? We fixed the architecture. Isolated validators. Human approval gates. TDD as a two-agent pipeline. We did the work. Then something broke. And nobody could explain what happened. The incident was recoverable — call it $340K in adverse selection, a bad afternoon on a $2B book. The problem was what came after. The CTO tried to reconstruct the agent decision chain. Three monitoring dashboards, a Slack thread, a Git log, a Confluence page that referenced a ticket that referenced another ticket. Four minutes later, he stopped. There was no single place that could show what each agent decided, in what order, and why. What signal passed between Design and Code. Whether Test flagged something that Deploy overrode, or whether the flag was never raised. Every agent had passed its own tests. Every agent had done exactly what it was told. The system failed anyway. 💡 This is the traceability problem. And almost nobody has solved it. UC Berkeley's MAST framework (arXiv 2503.13657, March 2025) analyzed 1,600+ execution traces across multi-agent deployments. Finding: the majority of failures come from inter-agent interactions, not individual agent limitations. The most common type — "reasoning-action mismatch" (FM-2.6) — 13.2% of all failures. The agent reasons correctly but acts on drifted context. That mismatch propagates silently across three or four stages before a human sees anything. Same validator bias problem I wrote about in my last post — but worse. Now you need to trace it backward through multiple deployment cycles. 🚫 Here is what makes this urgent. FINRA's December 2025 oversight report explicitly names multi-agent traceability as a supervisory concern: outcomes from multi-step agent reasoning are "difficult to trace or explain, complicating auditability." They expect firms to maintain "process reconstruction records" — show what your agents decided and why, in a format a regulator can review. EU AI Act Article 12 goes further: automatic event logging mandated for high-risk AI systems. Algorithmic trading qualifies. Enforcement begins August 2, 2026. Less than five months away. In my experience... approximately none of the CTOs running these pipelines can produce what FINRA is describing. They have observability. They do not have traceability. Those are very different things. I am building the infrastructure pattern that closes this gap. Architectural design choice at the agent boundary level, not a logging bolt-on after the fact. I will break down the architecture and cost math in my next post. If your multi-agent system produced a silent failure right now, could you reconstruct what each agent decided and why? 🤔 If the answer is "we would need to check"... that is the answer. #hft #aiagents #electronictrading #tradingtechnology

97

16 Comments

Applying Transformers to Financial Time Series In the previous blog post, we gave a very simple example of how traders can use self-attention transformers as a feature selection method: in this case, to select which previous returns of a stock to use for predictions or optimizations. To be precise, the transformer assigns weights on the different transformed features for downstream applications. In this post, we will discuss how traders can incorporate different feature series from this stock while adding a sense of time. ... Read more at https://lnkd.in/g8NRYAXY

214

5 Comments

PyQuant News highlights the Sortino Ratio as a valuable metric for assessing portfolio risk and reward, with particular emphasis on downside risk: https://lnkd.in/ezxq5zd4 #AlgoTrading

6

You’ve probably heard the terms OMS, EMS, and PMS thrown around, but what’s the difference between them? Databento’s latest guide breaks it down clearly as trading tech continues to evolve.

27

5 Comments

"Why it's more interesting to launch on Bittensor than start from scratch (Credit: @Axilo on X) Why TAO?  - Companies building on Bittensor are effectively subsidizing expensive compute by lifting on top of the network effects of Bittensor and its token, TAO as the primary incentive that everyone in the ecosystem competes for. - Most AI crypto protocols focus on deterministic outputs, rewarding only one type of result as the single source of truth. This results in one-dimensional incentive structures that are hard to scale and difficult to align with broader community participation. - Bittensor should be viewed not just as an AI protocol, but as an incentive protocol — a system where miners compete to produce subjective, high-value outputs across various subnets. - Because those outputs are subjective (e.g. language, vision, decision-making), the competition itself becomes dynamic, persistent, and difficult to game. - That creates a multi-layered economic game between miners, validators, and subnet owners with coordination, routing, and performance all incentivized natively - Many of these protocols are also hindered by high FDV funding rounds that front load investor returns and weaken the opportunity for meaningful, community driven growth. - Fairer launches like Bitcoin, Ethereum, and even Doge created stronger early ecosystems. Bittensor reintroduces that fairness with embedded incentive mechanisms that attract real contributors and incentives competition at every layer of the network. - As an investor excited about TAO, you're not betting on a single model or commodity. Instead on the collective intelligence / commodities of the network. - Every innovation from miners across every subnet is denominated in TAO, making it a meta-bet on decentralized AI productivity itself. - Instead of exposure to a single output, you're gaining exposure to the creation, distribution and competition for all outputs across the network. - As open-source AI models proliferate and centralized moats erode, Bittensor will become inevitable.. Offering open, incentive aligned coordination for AI development and deployment. - It's not just a bet on a network. it's a bet on the economic infrastructure layer of open-source AI. - What makes Bittensor especially compelling is how many existing crypto teams choose to launch subnets which proves that the economic design is working. I can see the same case for web2 companies coming to TAO to subsidize costs through miners. - It's simply more efficient to plug into TAO, align with existing miners and incentives, than to bootstrap a network from scratch. - Imagine if Midjourney, Stable Diffusion, or Hugging Face APIs were powered by miners competing to serve faster, cheaper, better results with the cost offset by TAO incentives. Whilst miners are hyper-competitive and have the incentives to do whatever it takes to outcompete their peers... That's a world I want to live in, and I believe that day will come."

12

4 Comments

At KERN., we replace speculative hope with mathematical certainty. Our proprietary MyQuant engine is designed to democratize quantitative intelligence, bringing hedge-fund-level stochastic modeling to the modern investor. The Methodology: 10,000 Alternate Universes We don’t rely on static price targets. Every KERN analysis utilizes a Monte Carlo engine to simulate 10,000 parallel price paths based on Geometric Brownian Motion. This high-fidelity approach allows us to calculate the Expected Value of a trade, that is the core metric of long-term profitability. Current Market Dispatch: Our recent head-to-head simulation reveals a stark contrast in risk-adjusted opportunities: SPY (S&P 500) $700 Call: Despite the benchmark’s stability, MyQuant identified a heavily negative EV of -$703.30. The target sits too far outside the high-probability "distribution meat," signaling a high-conviction "No-Trade" scenario. VDE (Vanguard Energy) $200 Call: Factoring in geopolitical catalysts in the Strait of Hormuz, the model identified a positive EV of +$578.45. VDE remains our superior tactical candidate for capital allocation. The KERN Advantage Successful execution is a function of managing probabilities. By prioritizing trades where the mathematical "wind" is at our back, KERN ensures that every entry is backed by institutional-grade validation. The App is in development, release date TBD. #KernQuant #InstitutionalIntelligence #QuantitativeFinance #RiskManagement #MarketAnalysis

5

A fantastic deep dive by William Arnold (NVIDIA) into Dynamo, an open-source system that rethinks high-performance LLM inference. If you're building or hosting models, this talk is a must-watch. Learn about smart cache routing, disaggregated serving, and real production-ready design. At AER Labs, we’re passionate about efficient AI + system-level innovation. Do subscribe to our channel for more such expert sessions. Our Batch V1 applications are now open. If you want to contribute to open-source AI research (part-time or full-time) or building a cool AI project, do check it out. Free and open to all. Fill this form to connect: https://lnkd.in/dJG8bXtt

12

In a world saturated with complex algorithms and opaque derivative flows, even imperfect signals from the physical economy can offer useful reference points. https://hubs.li/Q045DPXX0 Written by Ivan Illán, MHKSI, AIF®, CFS® of Aligne Wealth Advisors Investment Management (AWAIM®)

5

1 Comment

For those who prefer code over theory - complete Transformer implementation for financial time series: ✓ Custom FinancialDataset with configurable lookback windows ✓ Feature engineering: RSI, MA ratios, volatility, volume patterns ✓ Positional encoding implementation ✓ CLS token approach (BERT-style) for sequence aggregation ✓ Proper temporal train/test splits (critical!) ✓ Gradient clipping for stability ~430K parameters, runs on CPU or CUDA. All code tested and ready to run. 🔗 https://lnkd.in/eZK4Uc9r #Python #PyTorch #MachineLearning #TimeSeries

2

Today we released the whitepaper for the Open Cap Table Protocol (OCP). If you believe equity should move like code, this is the spec. OCP is the MIT-licensed, open source onchain infrastructure behind Fairmint. It turns equity into smart contracts, enforces compliance at the protocol level, and keeps cap tables accurate in real time. It’s not just a better experience. It’s a legal-grade system of record, live in production and fully aligned with the Open Cap Table Format (OCF) by the Open Cap Table Coalition The timing couldn’t be better. This week, SEC Chairman Paul S. Atkins said on-chain securities have “the potential to remodel aspects of the securities market.” Equity deserves real rails. OCP is the foundation. 🔗 PR: https://lnkd.in/ewcfr7sG

56

15 Comments

Gimme Credit cited in Quartz (10/6/2025): “Nvidia has indicated that demand is far outpacing supply, but it is not clear whether the company could have fulfilled an expanded order from OpenAI,” said Dave Novosel, Senior Bond Analyst at Gimme Credit. Read Full Article Here: https://hubs.la/Q03MyrkZ0 #gimmecredit #corporatebondresearch #inthenews

2