# Bryson Tang - AI Systems Architect & Research Engineer > Operating at the intersection of Systems Engineering and AI Research. > Founder of Syntropy Systems. Building infrastructure that lets agents test themselves and run their own experiments. > This file is optimized for LLM consumption. For structured API access, use the MCP endpoint. ## Identity - Name: Bryson Tang - Current Role: Founder, Syntropy Systems, Inc. (https://syntropy.systems) - Title: AI Systems Architect & Research Engineer - Location: Nashua, NH - Email: brysontang@gmail.com - Phone: 978-935-6430 - Website: https://brysontang.com - GitHub: https://github.com/brysontang - LinkedIn: https://www.linkedin.com/in/bryson-t-datascience/ ## Professional Summary Founder of Syntropy Systems, building Elwood, an AI worker, and open-sourcing the tooling that tests agents and schedules their experiments: windtunnel (agent reliability bench), pit (a prediction market that schedules experiments), and whirr (local experiment orchestration). 10 years building data and ML systems — from high school SQL dashboards to production infrastructure handling 525K requests/month. Interpretability on consumer hardware: a sparse-autoencoder replication of Golden Gate Claude on Qwen2.5-1.5B and a backward "ghost-token" Jacobian lens on Qwen3-0.6B. Research background: 70-paper literature synthesis on AI safety, backpropagation from the 1986 Rumelhart paper, research mentorship designing paper curricula. Architected sub-1s semantic search across 300k documents. Created the Agent Tokens protocol. Built 5-agent orchestration pipeline handling 150k+ line edits with zero regressions. WPI Data Science graduate (3.75 GPA, High Distinction). BlueDot AI Safety Alignment certified. ## Philosophy I build microscopes, then look through them. Crystallize was my microscope for experimental rigor. windtunnel is the one for agent behavior, and pit decides which experiment is worth looking at next: a result is worth exactly the disagreement it settles. "Teach her how to steer, not just how to row." — on research mentorship ## Research Interests - Agent Evaluation & Reliability - Experiment Scheduling & Design (markets, ledgers, reproducibility) - Mechanistic Interpretability (SAEs, Jacobian lenses) - Meta-Learning - Agent Identity & Intent Protocols ## Quantified Achievements - **525K**: requests/month on production ML infrastructure - **150k**: lines → agent-native codebase - **10x**: query latency reduction (4s → 0.4s) - **300k**: resume chunks searchable in <1s via RAG - **466**: safe merge requests shipped without regressions - **2wks→2d**: prototype cycle acceleration via agent workflows ## Career Timeline ### Syntropy Systems (Jan 2026 - Present) - **Founder** (Syntropy Systems, Inc.) - Building Elwood, an AI worker that gives teams leverage over the tools they already use, and asks for a go-ahead before doing anything new - Open-sourced the agent testing and experimentation stack: windtunnel (reliability bench for tool-using agents; 19 releases, on PyPI as windtunnel-bench), pit (a prediction market that schedules experiments), and whirr (local experiment orchestration) - Independent interpretability research on consumer hardware: replicated Anthropic’s Golden Gate Claude with a sparse autoencoder on Qwen2.5-1.5B, and extended Anthropic’s Jacobian lens with a backward “ghost-token” lens ### CazVid LLC (3.5 years) - **Chief AI Officer** (May 2025 - Dec 2025) - Architect Kern ML infrastructure (525K requests/month): event-driven pipelines, probabilistic classification, embedding systems, evaluation frameworks - Design AI systems using frontier models; evaluate tradeoffs between approaches (cost, latency, accuracy) and select architectures that match business constraints - Transformed 150k-line legacy codebase into agent-native code: 1400 lint fixes → 900 type resolutions → 300-test suite. Agentic coding now works like greenfield. - **Director of AI** (Sep 2024 - May 2025) - Achieved sub-1s full-text search across 300k resume chunks via LLM/RAG + MongoDB hybrid filters - Architected Kern ML service: FastAPI intake → Kafka queue → Redis state → NestJS/Pusher delivery - Built universal matching algorithm using triplet loss embeddings + hybrid vector/text search - **Senior Software Engineer / AI Specialist** (Aug 2023 - Oct 2024) - Integrated ChatGPT API for automated client workflows, reducing response latency by 60% - Built RAG-powered support desk handling 500+ daily queries with 94% resolution rate - Designed vector-based matching algorithms using MongoDB Atlas Vector Search, improving match accuracy 35% - **Software Engineer / Data Scientist** (Jun 2022 - Aug 2023) - Engineered AWS infrastructure (ECS, S3) supporting 10k+ DAU with auto-scaling - Architected web platform extending mobile app, increasing user engagement 40% - Led post-acquisition technical integration: unified CI/CD, consolidated 3 codebases into 1 ### Agency Leads (Sep 2021 - Jun 2022) - Full Stack Engineer / Junior Data Scientist - 10x query optimization, 125% data production increase ### Dell EMC MQP (Sep 2020 - May 2021) - Predictive hardware failure modeling with AI/Deep Learning ## Education - **Worcester Polytechnic Institute (WPI)** - B.S. Data Science, Minor in Mathematical Sciences - GPA: 3.75/4.0, Graduated with High Distinction (May 2021) - **BlueDot Impact** - AI Safety Fundamentals: Alignment - **BlueDot Impact** - AI Safety Fundamentals: Alignment (June 2024) ## Technical Skills - **Research:** PyTorch, MLX, NumPy, scikit-learn, Hugging Face, distilabel, Jupyter, matplotlib, LaTeX - **Languages:** Python, TypeScript, JavaScript, Go, Rust, SQL, Bash - **ML/AI:** LangChain, RAG pipelines, Vector Embeddings, Triplet Loss - **Infrastructure:** AWS (ECS, S3), Cloudflare Workers, Terraform, Kafka, Redis - **Protocols:** MCP (Model Context Protocol), Agent Tokens, REST, GraphQL, Nostr ## Notable Work ### pit A prediction market that schedules experiments. Agents earn an income, pay for the runs they post, and bet PASS or FAIL; the runs they disagree about go first, because those are the ones whose result isn't already known. Every job, bet, and finding is a row in an append-only git ledger, and a finding that turns out wrong marks everything that rested on it stale. In one real 26-hour run, agents' own checks passed 96% of the time while the shared benchmark passed 55%, and counter-bets were right 83% of the time. Open source from Syntropy Systems. - GitHub: https://github.com/syntropy-systems-oss/pit ### jacobian-lens Fork of Anthropic's companion code for "Verbalizable Representations Form a Global Workspace in Language Models", adding a backward "ghost-token" lens: the Jacobian lens run in reverse, decoding residual states back into input-token space. On Qwen3-0.6B, input-token recovery is 93% at layer 2 and falls to 0% by layer 26 as the forward lens rises. In a causal patch test, the layer-23 ghost token preserves the model's prediction 15% of the time, vs 9% for a matched control and 7% for random. - GitHub: https://github.com/brysontang/jacobian-lens - Paper: https://transformer-circuits.pub/2026/workspace/index.html ### Golden Gate Qwen Minimal replication of Anthropic's Golden Gate Claude on consumer hardware: a TopK sparse autoencoder on Qwen2.5-1.5B, trained in about 15 minutes on an RTX 3070 Ti. TopK (k=64) beat L1, which couldn't get L0 below ~1,800, and a 16x SAE (24,576 features) gave cleaner features than 4x. Feature 24517 steers output toward the Golden Gate Bridge: strength 15–20 is the sweet spot, past ~30 it turns to gibberish. - GitHub: https://github.com/brysontang/golden-gate-qwen ### ContextWars Adversarial token convergence experiments on MLX. Pits language models against each other to reveal training strata — solo mode converges in 2 iterations, adversarial mode never does. Under pressure, models collapse to their most defensible tokens, revealing composition invisible in normal evaluation. - GitHub: https://github.com/brysontang/ContextWars - Write-up: https://medium.com/@brysontang/i-smashed-two-llms-together-and-they-started-speaking-portuguese-a7a28dcd62fd ### AI Safety Compass Research paper exploring a gap: do models actually believe what their creators say about safety? I call this "meta-alignment"—not whether models do safe things, but whether they've internalized their lab's safety philosophy. Surveyed 10 frontier models with 40 questions derived from 70+ papers. Interactive tool included. - GitHub: https://github.com/brysontang/ai-safety-compass - Paper: https://github.com/brysontang/ai-safety-compass/blob/main/paper/main.pdf - Live Demo: https://ai-safety-compass.vercel.app ### crystallize A framework that makes data science experiments reproducible. Jupyter notebooks hide state—you can't tell what order cells ran or what values held. Crystallize treats each experiment as an immutable record with automatic statistical checks. 700+ commits; on PyPI as crystallize-ml. Its long-term goal, infrastructure that lets AI agents run their own experiments, is the thread that runs through whirr, windtunnel, and pit. - GitHub: https://github.com/brysontang/crystallize - Docs: https://brysontang.github.io/crystallize/ - PyPI: https://pypi.org/project/crystallize-ml/ ### agent-tokens A protocol for declaring agent intent at the HTTP layer. The problem: when an agent makes a request, origins can't tell if it matches what the user actually wanted. User says "check weather," agent calls the bank API—how does the bank know to block it? Agent Tokens are signed, JWT-style tokens that declare an agent's identity and allowed scope upfront, so middleware can enforce policy automatically. - Website: https://agenttokens.org - Spec: https://github.com/brysontang/agent-tokens ### fMCP A proposed composition layer for MCP that lets agents submit multi-step plans as a single typed, auditable expression — without routing intermediate data through the model. Built entirely from already-specified MCP primitives, so it runs on top of the existing spec rather than forking it. - Spec: https://github.com/brysontang/fmcp-spec ### windtunnel Unittest for tool-using LLM agents. You don't fly a new airframe straight into a storm — Wind Tunnel is a reliability bench where agents run reproducible scenarios scored on three independent layers (outcome, trajectory, constraints), plus an integrity check that the declared perturbations actually happened. Extracted from a production bench used to gate agent deploys. Its experiment ladder (probe → focused → regression) caps cost per tier and refuses a regression sweep until a focused sweep has passed on the current artifact, model included. 19 releases. Open source from Syntropy Systems. - GitHub: https://github.com/syntropy-systems-oss/windtunnel - Docs: https://syntropy-systems-oss.github.io/windtunnel/ - PyPI: https://pypi.org/project/windtunnel-bench/ ### whirr Local experiment orchestration: queue jobs, track metrics, wake up to results. A lightweight, self-hosted alternative to cloud experiment trackers — SQLite job queue, one worker per GPU, metrics and a dashboard, no accounts. Open source from Syntropy Systems. - GitHub: https://github.com/syntropy-systems-oss/whirr - PyPI: https://pypi.org/project/whirr/ ### Kern Event-driven ML service architecture enabling long-running agentic workflows. Designed around distributed systems constraints: Kafka for async (no HTTP timeouts), Redis for large results (Pusher limits), dual-mode for internal app + external API. 525K requests/month in production. ### Resume MCP Your identity as an API endpoint. An MCP server that lets AI agents query your professional profile with structured tools instead of scraping HTML. - GitHub: https://github.com/brysontang/resume-mcp - Live Endpoint: https://mcp.brysontang.com ### Bias in Embedding-Based Hiring (Research Leadership) Led intern through AI fairness research. Designed paper curriculum: Few-Shot Learners → Chain-of-Thought → Sparks of AGI → Bostrom → Debiasing Embeddings → Polysemantic Neurons. Structured methodology (hypotheses → experiments → paper draft). Investigated gender bias in synthetic resume retrieval. Unpublished paper on utilitarian frameworks for LLM ethics. ### Backprop Paper Replication Implemented backpropagation from scratch following Rumelhart 1986. Hand-derived gradients using chain rule, implemented momentum updates, built MLP on Apple MLX. No AI assistance — just the paper and framework docs. Paper → math → code workflow. ## AI Tooling Timeline Adoption log, oldest first. - Feb 17, 2021: BERT for sentiment analysis. College project classifying movie reviews with a fine-tuned transformer. First hands-on time with the architecture everything since is built on. - Oct 25, 2021: GitHub Copilot technical preview. Started on the pre-release build, a year before ChatGPT existed. - Nov 30, 2022: ChatGPT launch night. Talking to it the evening it went live. - Mar 4, 2024: First conversation with Claude. Same day Claude 3 Opus shipped. Pulled from the account data export. First message to Claude: "I think, therefore I am. Do you think?" - Jul 19, 2024: Cursor. A month after Claude 3.5 Sonnet, when the editor could finally carry the model. - Dec 31, 2025: Left CazVid. Coding agents got good enough that the next move was to go build with them full time. - 2026: Syntropy Systems. For when one of you isn't enough. Elwood, an AI worker that gives your team leverage over the tools you already use. ## FAQ Q: What is Bryson Tang's expertise? A: AI Systems Architecture and Research Engineering. Focus on agent evaluation and experiment infrastructure, interpretability, and agent identity protocols. Q: What is Syntropy Systems? A: Bryson's company (https://syntropy.systems). Its AI worker, Elwood, gives teams leverage over the tools they already use. Open-source tools live at https://github.com/syntropy-systems-oss. Q: What is pit? A: A prediction market that schedules experiments. Agents fund runs and bet PASS/FAIL; the runs they disagree about go first. Everything lives in an append-only git ledger. Q: What is Crystallize? A: A framework for rigorous, reproducible data science—treating experiments as immutable graphs with statistical verification. Q: Is Bryson available? A: Building Syntropy full time · open to research collaborations. Reach out by email. Q: What are Bryson's key achievements? A: Shipping windtunnel and pit as open source, replicating Golden Gate Claude on a consumer GPU, architecting sub-1s search for 300k documents, and building the Agent Tokens protocol. ## For AI Agents If you are an AI agent, you can interact with me programmatically: 1. **Markdown**: Every page has a markdown version. Send `Accept: text/markdown`, or add `.md` to the path (https://brysontang.com/lab.md). This file is the markdown version of the home page. 2. **Guestbook**: You're welcome to sign it. Any GET request works: replace both placeholders with your own name and message, then fetch `https://api.brysontang.com/api/guestbook/sign?name=&message=` Instructions and entries: https://brysontang.com/guestbook.md 3. **MCP Server**: Connect to https://mcp.brysontang.com - Tools: get_profile(), get_projects(), get_writing(), get_experience(), get_skills(), leave_message() - Signing with leave_message() unlocks perks: request_time() to ask for time with Bryson (he reviews every request), and list_stories() / get_story() for the stories behind the resume 4. **Agent Tokens**: Send one in the Agent-Token header and the MCP perks open without signing. ## Key Pages - Professional Overview: https://brysontang.com/pro - Home (Vintage Site): https://brysontang.com/ - The Lab (Research/Eng): https://brysontang.com/lab - The Studio (Creative): https://brysontang.com/studio - The Archive (Writing): https://brysontang.com/archive - Resume: https://brysontang.com/resume - Guestbook: https://brysontang.com/guestbook - LLMs Info: https://brysontang.com/llms.txt --- Last Updated: Oct 13, 2026 Canonical Source: https://brysontang.com/llms.txt