agents-ops

A personal platform of AI agents and AI products, built by Sergio Gabriel Fruto to answer one question: what does it take to move applied AI from demo to something that runs with real stakes?

▶ Live demo: Biomech Lab — upload a yoga pose, get it recognized and scored from joint angles measured entirely in your browser.

---

What's here

Project	What it does	Status	Stack
Biomech Lab	Recognizes Warrior II, Tree and Downward Dog in a photo and scores posture from measured joint angles. Images never leave the browser.	Live · biomech-lab.vercel.app	Next.js 16 · TypeScript · MediaPipe Pose · Vitest · Vercel
Polymarket agent	Scores prediction markets, estimates true probability, sizes with half-Kelly, places and exits CLOB orders.	Active (dry-run by default, live mode available)	Python · Polymarket CLOB · Claude Haiku · SQLite
Dota agent	Predicts pro Dota 2 match outcomes from Elo and history; simulates Kelly-sized bets.	Simulation only	Python · OpenDota API · SQLite
Job Hunter	Scores job postings against a candidate profile, writes tailored cover notes, syncs roles to the dashboard.	Active	Python · Claude API · SQLite
The Analyst	Daily intelligence brief: CVEs (NIST NVD, CISA KEV), AWS security bulletins, OSINT signals (OpenSky, GDELT), synthesized into one report.	Active (daily)	Python · Claude API · Next.js
Silicon Intel	Semiconductor supply-chain and hardware-security terminal; Claude writes a daily brief from industry, CVE and materials feeds.	Paused	Python · Claude Sonnet · RSS feeds
Finance agent	Personal finance tracker (statements, goals, net worth).	Paused, local use	Python · SQLite
Solaris	Coordinator + dashboard that runs the agents in dependency order and writes a daily operations log.	Active	Python · Flask · Next.js 16 · shadcn/ui

---

Biomech Lab

The newest project, and the one you can click.

• Pipeline: MediaPipe detects 33 body landmarks in the browser → pure TypeScript modules convert them to pixel space, infer the camera view (front/side) and framing, classify the pose by rules, and score each check (e.g. "front knee 90° ± 10°") with partial credit.
• Honest scoring: a readiness gate refuses to show a number when the view is wrong, the body is badly framed, more than one person is in frame, or less than 70% of the checks can be measured — the visitor gets a specific prompt instead.
• Privacy: images stay on the device. MediaPipe ships a built-in usage logger with no opt-out; a connect-src Content-Security-Policy blocks it, verified in production.
• Tested: 83 unit and golden tests, including scores from real sample photos and mirrored-input invariance.
• Next: a streamed Claude coach for feedback on the weakest checks, and live camera mode.

Details: biomech-lab/README.md · design spec

---

Solaris — how the agents run together

agents-ops/
├── solaris/              ← Coordinator + dashboard
│   ├── coordinator.py    ← DAG orchestrator
│   ├── roadmap_updater.py
│   ├── daily_log.py      ← Claude synthesis
│   ├── SOLARIS.md        ← Coordinator system prompt
│   └── web-next/         ← Next.js 16 dashboard
│
├── polymarket-agent/     ← Prediction market trading
├── dota-agent/           ← Esports match prediction (simulation)
├── job-hunter-agent/     ← Job opportunity pipeline
├── analyst-agent/        ← Security & OSINT intelligence brief
├── stock-agent/          ← Silicon Intel (paused)
├── finance-agent/        ← Personal finance (paused)
└── biomech-lab/          ← Standalone web app (not orchestrated)

Each agent declares a roadmap.yaml — tasks, dependencies, schedule and outputs. The coordinator reads all of them at startup, builds a dependency graph, runs agents in topological order, and writes a Claude-generated daily operations log.

Startup
  └── Read all roadmap.yaml files
  └── Build dependency DAG → topological sort

For each agent (in order):
  └── Inject: tasks_today + memory + upstream outputs
  └── Run agent subprocess
  └── Collect outputs
  └── roadmap_updater.py → mark tasks done / blocked

End of day:
  └── Claude call → daily_logs/YYYY-MM-DD.md
       ├── Agent status table
       ├── Blockers
       ├── Cross-agent patterns & signals
       └── Prioritized suggestions

Coordinator (coordinator.py)

• Discovers agents by scanning */roadmap.yaml
• Builds a dependency DAG and runs Kahn's topological sort
• Dispatches each agent as a subprocess, injecting context via the COORDINATOR_CONTEXT env var
• Agents report back via a JSON sentinel line on stdout: {"__coordinator_outputs__": {...}}
• Updates each roadmap.yaml after the run (task status, memory, last output)

Daily log (daily_log.py + SOLARIS.md)

• One Claude call at the end of each run
• Input: run results, current roadmap state for all agents, yesterday's log
• Output: structured Markdown — agent status table, blockers, cross-agent signals, suggestions

Dashboard (web-next/)

• Next.js 16 server components, shadcn/ui, Tailwind; proxies /api/* to the Flask backend
• Sections: agent health, market data (BTC, stocks, Fear & Greed), job pipeline, Polymarket bets, Dota analytics

---

Polymarket agent

• Fetches active markets from the Gamma API and scores them on a composite signal: probability sweet-spot, volume, bid-ask spread, price stability, days to expiry
• Derives true probability from domain signals: crypto volatility (CoinGecko), sports odds (The Odds API), and Claude Haiku for everything else
• Sizes positions with half-Kelly against a bankroll with exposure caps
• In live mode, places limit orders on the Polymarket CLOB (L2 auth, Polygon) and exits autonomously on stop-loss / take-profit
• Dry-run by default (DRY_RUN=true): every decision is recorded without placing orders

---

Tech stack

Layer	Technologies
Languages	Python 3.11 · TypeScript (strict)
AI / LLM	Claude API (Anthropic) — signals, synthesis, cover notes, intelligence briefs
Computer vision	MediaPipe Pose Landmarker (WASM/WebGL, in-browser)
Frontend	Next.js 16 · React 19 · Tailwind · shadcn/ui · Canvas API
Backend	Flask · SQLite (per-agent DBs)
Trading & data	Polymarket CLOB · py-clob-client · CoinGecko · The Odds API · yfinance · OpenDota
Orchestration	Custom DAG coordinator (Kahn's algorithm) · schedule · per-agent roadmap.yaml
Testing & deploy	Vitest · Vercel

---

Running

Each project is independent.

# Biomech Lab (Node 22.12+)
cd biomech-lab && npm install && npm run dev      # http://localhost:3000
npm test                                           # 83 tests

# Polymarket agent (dry-run by default)
cd polymarket-agent
pip install -r requirements.txt
cp .env.example .env   # fill in API keys
python main.py

# Solaris dashboard
cd solaris && pip install -r requirements.txt && WEB_PORT=5002 python main.py
cd solaris/web-next && npm install && npm run dev

# Coordinator (runs all enabled agents)
cd solaris && python coordinator.py

Live trading requires Polymarket credentials (DRY_RUN=false + POLY_PRIVATE_KEY); run python live_setup.py once to generate CLOB API keys.

---

Design principles

• Each agent is sovereign — runs standalone, no hard dependency on the coordinator
• The coordinator is additive — agents don't need to know they're being orchestrated
• LLMs where they earn their cost — Claude handles ambiguous signals and synthesis; deterministic code handles measurement and scoring
• Real stakes, explicit switches — live trading is opt-in, with stop-loss and take-profit exits
• Privacy by default — Biomech Lab keeps images on the device and blocks third-party telemetry
• Roadmap as runtime state — roadmap.yaml doubles as task tracker and agent memory