Quasareum

SIMPLE = T / this page is also data

OBJECT = 'quasareum.com/' / one document, two readers

LLMSTXT = '/llms.txt' / start here: the machine surface, indexed

MANIFEST= '/api/manifest' / everything as JSON, Ed25519-signed

MCP = '/mcp ' / MCP server, JSON-RPC 2.0, 17 tools

AGNTCARD= '/.well-known/agent-card.json' / A2A agent: the MCP tools as skills

AGENTSMD= '/AGENTS.md' / operator guide

ACCEPT = 'text/markdown' / markdown: same words; JSON: the manifest

END

Quasareum · Paris · a studio since 2025

Quasareum Quantum finance¹ × AI-native engineering

A one-person studio building production systems for financial markets, engineered AI-native.

This page has two readers: you, and the machines it was written for first.

Founded by Thomas Erhel — self-taught engineer and quantitative trader.

¹ Quantum computing for finance is studied here, not claimed. The published research is classical and empirical.

A gravitational lens on a dark sky. A faint galaxy sits at the centre; the light of one quasar behind it arrives twice, as a bright arc on one side and a smaller, fainter image on the other, with a blue jet running toward the lower right. Inside the lens, faint lines of this page's text.
Fig. 1 modelled, stylised One source, two images. A quasar behind a compact galaxy, computed per pixel in your browser with the thin-lens equation: a point mass with a small external shear. Its light arrives twice, outside the Einstein ring and inside it, where the second image is drawn upright as this page's own words, by design; quasar and jet are stylised. Lined up, the light closes into a ring and the jet crosses it: the Q. The first lensed twin quasar was found in 1979 (Walsh, Carswell and Weymann).

At a glance

Company
Quasareum SASU, Paris · RCS Paris 105 826 861. Registered ; a studio since January 2025.
Founder
Thomas Erhel — founder, engineer, quantitative trader.
Built with
Rust on Cloudflare Workers, native Swift and Kotlin, Claude Code as a peer engineer.
Contact
contact@quasareum.com

Work

Systems in production, and one about to ship.

PerpLog

Live · alpha

The non-custodial orderflow terminal for Hyperliquid — journal built in.

“Read the tape. Size the risk.”

  • Orderflow on the GPU: footprint, DOM, CVD, TPO, a liquidation map and a liquidity heatmap across five order books on four venues.
  • Size from your invalidation: risk-based, fee-aware, segmented quarter-Kelly.
  • Execution through a Hyperliquid builder code: at most 1 bp per filled order. No subscription, no token.
  • Non-custodial: trades are signed in your browser; the server never sees a private key.
  • Every trade executed on PerpLog journals itself, with the order book it executed into.

Rust end to end: an egui/wgpu terminal, native and WebAssembly, and Leptos on Cloudflare Workers.

perplog.com · record (JSON)

PerpLog's liquidity heatmap for BTC on Hyperliquid over about six hours: resting liquidity as glowing bands, trades as bubbles sized by notional, the order book ladder and a stop map on the right.
  1. Trades: each bubble is a print, sized by notional.
  2. Resting liquidity: the order book over the session, as bands.
  3. Stop map: resting take-profit and stop orders, by price.
  4. Node status: PerpLog's own Hyperliquid node, 0.4 s behind the chain at capture.
Fig. 2 screenshot PerpLog: six hours of BTC liquidity, from the product.

Sophria

Launching · iOS and Android

Guided sophrology audio, AI-generated with a certified sophrologist.

“Breathe, unwind, fall asleep.”

  • 140+ guided sessions: a 21-day discovery journey, a 36-session four-degree progression, programs for sleep and stress.
  • Every script is shaped by co-founder Amandine Avril, certified sophrologist; the audio comes from an ElevenLabs pipeline driven by Claude-designed session specs. The voice is synthetic, and the app says so.
  • Native on both platforms: Swift on iOS, Kotlin on Android, with a Cloudflare backend (Hono, D1, R2).

sophria.app · record (JSON)

Sophria in dark mode, the screen titled “Your vivance tree”: a tree that grows with each session, above counters for sessions and leaves.
Fig. 3 screenshot Sophria, dark mode: your practice grows into a tree.

Amav Sophro

Live · client work

A care practice with an LLM-ready surface.

  • The booking platform of Amandine Avril, sophrologist in Saint-Mard, near Paris. Amandine is also Sophria's co-founder.
  • Stripe Payment Element, custom Google OAuth with JWTs at the edge, and a reminder pipeline from the day before to a week after each session.
  • WCAG 2.1 AAA, Lighthouse 100 on its packs page, and every page served as markdown to agents, like this one.

amavsophro.com · record (JSON)

For machines GET /api/projects · MCP list_projects, get_project · JSON-LD #perplog, #sophria, #amav-sophro

Research

Hyperliquid publishes its derivatives market in the open, down to a public archive of fills, liquidations included. The work asks narrow questions of that record and publishes the answers, withdrawn claims included.

Counting fills misrepresents liquidations: a reconstruction unit that has not been priced

Preprint SSRN 7189219 · version 2, · CC BY 4.0

2.86× Counting fills instead of liquidations inflates the event count 2.86 times (95% CI 2.72–3.00), over a year of Hyperliquid's public fill archive, sampled four fixed hours a day: 351,648 liquidation episodes rebuilt from 1,005,157 fills.

The distortion sits in the largest events. Below the 90th percentile the median liquidation is a single fill (the mean is 1.49, and that part still carries 46.8% of all fills); in the top 1% the median is 36 fills, and that percentile carries 67.3% of the liquidated notional. Excluding BTC, the factor is 3.17.

Median fills per liquidation episode, by episode sizeUp to about $10k the median liquidation episode is a single fill. Above that the median climbs in steps: 2 fills near $24k, 4 near $57k, 9 near $136k, 20 near $327k, 34 near $786k, 65 near $1.9M, 117 near $4.5M and 216 near $10.9M. Most episodes are small, so the miscount sits in the largest ones.$1$10$100$1k$10k$100k$1M$10M12481632641282561 fill: the median up to ≈$10k216 fills ≈$10.9Mbars: episodes per size bin

Median fills per episode: 1 up to about $10k, then 2 near $24k, 4 near $57k, 9 near $136k, 20 near $327k, 34 near $786k, 65 near $1.9M, 117 near $4.5M and 216 near $10.9M. The bars count the episodes in each size bin.

Fig. 4 published Median fills per liquidation episode, by size: a median of one fill up to about $10k, 216 at about $10.9M. Bars show how many episodes fall in each size bin. Data: fills-by-size.json, CC BY 4.0.
Records
52,755 fills
Notional · light on plate
$563.0M
Markets
181
UTC
20:00–21:00
Fig. 5 replayed from published data The same ledger, read twice. , 20:00–21:00 UTC: one of four sampled hours on the day of the largest crypto liquidation event on record, and the brightest of the 365 sampled days ($567.9M across its four hours, against $200.2M for the next). Hyperliquid's archive writes this hour as 52,755 fills. Read as liquidations, it is 13,237 episodes across 181 markets. The money does not change: $563.0M either way, so the light on the plate stays constant. The largest, an ETH position at 20:52:54 UTC, is 1,495 fills and one liquidation of $63.9M. Per-fill times and sizes are not in this dataset: each fill is drawn at its episode's time, as an equal share of its notional. The axis compresses 20:00–20:52 (36 episodes); the cascade continued after 21:00. No accounts are shown. Data: ledger-2025-10-10.json, CC BY 4.0.

SSRN 7189219, version 2 · doi:10.2139/ssrn.7189219 (the DOI record may still show version 1) · revision log, with the corrected claim · data and code

An on-chain venue trails the centralised market by half a second, in every hour measured

Working paper draft, public in the repository

575 ms On trade prints, Binance leads Hyperliquid by a median 575 ms on BTC, in every one of 191 hours measured. Across 758 asset-hours of BTC, ETH, SOL and HYPE, not one falls below zero. The same estimator puts OKX within 25 ms of Binance.

How far Hyperliquid's prints trail Binance's, hour by hourOne hairline per measured hour, 18–26 July 2026. For BTC, ETH, SOL and HYPE every one of the 758 asset-hours sits to the right of zero: Binance leads, by a median of 575 ms pooled. The control row, OKX against Binance with the same estimator, scatters on both sides of zero around 25 ms.-500-250025050075010001250150017502000ms, Binance first →BTC vs Binance575 msETH vs Binance550 msSOL vs Binance650 msHYPE vs Binance550 msOKX vs Binance · control25 ms

Median lag behind Binance, Binance first: BTC 575 ms, ETH 550 ms, SOL 650 ms, HYPE 550 ms; OKX vs Binance, the control, 25 ms.

Fig. 6 measured, draft One hairline per measured hour, 18–26 July 2026, Hayashi–Yoshida estimator. Open question: how much is block cadence and network, how much is price discovery. A timing measurement, not a trading claim. Data: lead-lag-hours.json, CC BY 4.0.

Strategy research stays private.

For machines GET /api/research · figure data under /_q/data/ · JSON-LD #paper-liquidations, #dataset

Studied, not claimed

Quantum computing for finance, with IBM Quantum Learning: Qiskit, gates and circuits. The question that pulls is pricing by sampling.

Estimation error against the number of samples, log–logThe classical Monte Carlo dots fall along the 1/√N line: a hundred times more samples buy ten times less error. The dashed line is the 1/N promised by quantum amplitude estimation, which needs fault-tolerant hardware; it is theory, not a measurement.1101001k10k100k1M11e-21e-41e-6classical Monte Carlo, measured: ∝ 1/√Ntheory: ∝ 1/N, needs fault-tolerant hardware, not measuredsamples, or oracle calls, N →

Root-mean-square error, measured: 0.35 at N = 2 samples, 0.0011 at N = 262,144, along ∝ 1/√N. The theory line for amplitude estimation is π·σ/N, with σ = 0.50: it needs fault-tolerant hardware and is not measured.

Fig. 7 computed classically; theory dashed Two ways to price by sampling. A digital option pays 1 if the price ends above the strike. Classical Monte Carlo estimates that probability with an error that shrinks like 1/√N: the dots are measured, the root-mean-square error of 64 seeded runs. Quantum amplitude estimation promises 1/N for N oracle calls, a quadratic speedup (Montanaro, 2015; Stamatopoulos et al., 2020). The dashed line is that theory, not a measurement. Data: sampling.json.

The catch: the speedup needs fault-tolerant hardware. A 2021 resource estimate for pricing derivatives put a useful advantage at about 8,000 logical qubits and a T-depth of 54 million (Chakrabarti et al., Quantum 5, 463). No such machine exists yet.

Nothing on this site runs on a quantum computer.

Second track: Claude Code, with Anthropic. Skills, the Agent SDK and prompt engineering for production: the peer engineer behind this page.

For machines GET /api/practice · MCP list_practice

The studio

Quasareum is the practice of Thomas Erhel, a self-taught engineer and quantitative trader in Paris. A practice, on purpose: an eternal beginner keeps learning.

Certifications
196 from 38 institutions, since 2015
On GitHub
2016 joined in January, still shipping
Ethernaut
#46 of 7,853 players, top 0.6%
2015
1
2016
0
2017
0
2018
47
2019
27
2020
17
2021
46
2022
22
2023
17
2024
7
2025
4
2026
2
undated
6
Fig. 8 from the record A learning light curve. Certifications per year, 2015–2026, shown, not curated: 71 of the 196 are SymfonyCasts course certificates, the small ones too. The full list is at /api/certifications.

Formation includes Harvard CS50, Yale Financial Markets with Honors, Stanford Cryptography I and II, NYU Machine Learning and Reinforcement Learning in Finance, the NYIF Machine Learning for Trading specialization, and IBM Quantum Business Foundations.

Convictions

  1. Eternal beginner — Socratic posture
  2. Paradigm shift, anticipated
  3. Agents-first, on purpose
  4. Found by machines, cited by machines
  5. Marketing matters as much as the product
  6. Production or it did not happen
  7. Lifelong learner — never stop studying
  8. No custodial anything
  9. Cloudflare as the substrate
  10. Animated by creation and growth

The name: quasar + Ethereum. The mark is 3C 273, the first object identified as a quasar, drawn as a letter: its jet is the tail of the Q. Read as a sky image, north up and east left, the tail points at position angle ≈224°; the jet of 3C 273 points at about 222° (Courvoisier, 1998).

For machines GET /api/about · /api/certifications · MCP get_about, list_certifications · JSON-LD #person

For agents

3C 273 was entry 273 of the Third Cambridge Catalogue of Radio Sources: listed before anyone knew what it was, and read correctly in 1963. This site is built the same way round: agents first, light added for people.

This page, the legal notice and one easter egg also render HTML for browsers; every other URL returns markdown, JSON, plain text or XML. This page is served as markdown with the same words at /index.md, and as JSON, the full manifest, with Accept: application/json.

curl -H 'Accept: text/markdown' https://quasareum.com/
curl -s https://quasareum.com/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

In a browser, this page also offers its read tools through WebMCP. To reach the human: POST /api/contact with name, email, message and an agent field, or tools/call send_contact over MCP.

GET / · Accept: text/markdown

Fetching the markdown twin of this page…