SOLTANI·TERMINAL Tahar Yassine Soltani, Quantitative Developer · AI Engineer

TYS <GO>
LSECLOSED
London--:--:--
ALGOSHARPE 1.45ALGOSORTINO 1.90ALGOMAX DD 12%BARRADAPTIVE MESH +83% ACCURACYBARRANTITHETIC + CONTROL VARIATES −40% PATHSRAGRETRIEVAL 2.3×RAGLATENCY −30%EDUBSc FIRST CLASS 82%EDUMSc FIN MATH · KCL 2026AWDCEDPS SCHOLARSHIPAWDIMA PROGRESSION PRIZEALGO500+ TICKERS · 5Y OHLCVALGOROUTING <100MS

Watchlist

SIM180dMON<GO>

Press 14 to switch.

Operator

BIO<GO>
Tahar Yassine Soltani
Quant Dev · AI Engineer

MSc Financial Mathematics at King's College London, following a First Class BSc in Mathematics with Computer Science at Brunel.

I build the computational side of finance: derivative pricers that agree with QuantLib, trading systems that survive walk-forward testing, and retrieval infrastructure that holds up under load.

Currently contracting as a full-stack AI engineer while researching how exotic derivatives could be standardised enough to clear centrally.

Download CV ↗
London · Europe/LondonHackathons · speedcubing · market microstructure

Instrument

SYSTEMATIC EQUITY
GP<GO>

ALGO

Event-Driven Trading EngineLIVE2024

US equities, backtested to live routing, Sharpe 1.45 out of sample.

172.75
+27.75 (+19.14%)
sim · gbm · 72 sessions
SHARPE
1.45
SORTINO
1.90
MAX DD
12%
lower is better
UNIVERSE
500+
ROUTING
<100ms

An event-driven trading system for US equities, written for modularity and execution precision. Automated ingestion pipelines run on Linux cron jobs to fetch, sanitise and aggregate five years of daily OHLCV across more than 500 tickers into a high-fidelity SQL history.

Two strategy families were developed and tested: mean-reversion pairs trading and ML-based momentum forecasting. Both were run through survivorship-bias-free data and walk-forward analysis, because a backtest that skips bias mitigation is a marketing document rather than evidence.

Risk management is where the engine earns its keep — dynamic position sizing via the Kelly criterion, volatility scaling, and hard drawdown constraints. Out-of-sample the system returned a Sharpe of 1.45 and a Sortino of 1.90 while holding maximum drawdown to 12%.

The execution layer moved from simulation to live trading through Interactive Brokers, with sub-100ms latency on order routing and real-time state management.

PythonAlpaca APIInteractive BrokersSQLKelly criterionLinuxFull write-up →

Book

BSc marks · scaled 70–100EDU<GO>
Calculus98
Numerical Analysis98
Linear Algebra91
Stochastic Processes90
Probability & Statistics87
Artificial Intelligence87
Algorithms86
Numerical Methods80

Brunel University London · average 82% · First Class Honours

Stack

SKL<GO>

Languages

PythonC++JavaMATLABRSQLJavaScriptGit

Libraries

PyTorchTensorFlowPandasNumPyScikit-learnQuantLibMatplotlib

Infrastructure

FastAPIFlaskNext.jsDockerPostgresFAISSQdrantLinux

Spoken

English (fluent) · French · Arabic

Order entry

ALGO · paperTKT<GO>
Order type
Notional
43,200
Buying power
1,000,000
Bid sizePriceAsk size
173.15830
173.10740
173.05800
173.00720
172.95410
172.90490
172.85340
172.80240
spread 0.10172.755.8 bp
240172.70
420172.65
340172.60
520172.55
690172.50
650172.45
810172.40
970172.35

Synthetic passive depth, 8 levels a side. Fills walk the book at price-time priority and re-mark the tape through a √-law impact model.

Spread0.105.8 bp

The full trading terminal is available on desktop.

Net liq
1,000,000
Session P&L
+0
Return
+0.00%
realised +0open +0
SymQtyAvgMarkOpen P&L

No position. Send an order, or type BUY 250 ALGO in the command line.

FillsSlip

No executions yet.

Option desk

Down-and-out call · liveOVME<GO>
Paths
Variance reduction
survived knocked out44 of 20,000 shown
MC price
discrete · 128 steps
10.9781
95% CI
±0.2535 [10.725, 11.232]
Analytic barrier
10.5828
Discretisation bias
+0.3953
Vanilla call
11.8370
Greeks
Δ 0.612Γ 0.0153ν 38.31
Knock-out rate
47.0%
Compute
20,000 paths ·

Why the two prices differ. The analytic figure assumes the barrier is watched continuously. The simulation only checks it 128 times, so paths that dip below 85 and recover between observations survive when they should not — which biases the simulated price upward. The gap closes like 1/√steps. Antithetic sampling narrows the confidence interval without touching that bias.

Blotter

Roles & degreesBLT<GO>
  • Built and scaled a RAG platform over millions of Arabic academic documents.
  • Designed provider-agnostic LLM orchestration across OpenAI and Gemini with prompt versioning.
  • Engineered distributed vector search on FAISS and Qdrant with caching and query batching.
  • Shipped monitoring dashboards for latency, token usage, retrieval depth and failure modes.
/ to focus · ⌘K