ML ENGINEER FOR QUANTITATIVE TRADING — PREPARATION ONLY
Public demo: https://pz5l45g4wb77gg7q6idoazwaeq0eqvxn.lambda-url.us-east-1.on.aws/

Ingyu has built an AI-assisted Kalshi trade-tape quoting research lab using 989,895 public trade prints from 48 sampled KXBTC15M markets. The system models contract states throughout their life, compares logistic and gradient-boosted replay fill models, carries inventory, handles tick sizes and partial fills, and exposes execution assumptions. It provides a reproducible quoting API, chronological evaluation, fee/queue/latency sensitivities and a delayed-feedback LinUCB comparator.

On 11 held-out markets across two days, gradient boosting achieved fill-label Brier 0.2236 versus logistic 0.2525 and constant 0.2473. These are hypothetical fill labels inferred from actual trades. The learned value policy lost $24.64 in strict-through replay, versus $4.44 for the validation-selected fixed policy; no improved trading profitability is established. The anchored spot-digital fair-value approximation scored Brier 0.1797, but continuous BRTI basis and exact 60-second settlement averaging remain unestablished.

For the client, Ingyu would replace replay labels with their timestamped quote/order/fill/cancel history, reconstruct queue and latency effects, calibrate fill quantity and adverse-selection estimates, and validate policies on additional untouched market days. Deliverables would include audited data transformations, model artifacts, event-driven replay, baseline comparisons, operational quote-service controls, and a documented go/no-go decision based on net results.

Matcher recommendation: Ingyu offers an inspectable implementation of empirical trade-tape modeling, chronological evaluation and quoting failure analysis. His AWS demo makes the code, measured results and limitations directly reviewable. This evidence supports research/implementation work, but not a claim of a profitable production market-making engine.

Historical experience remains as documented in the existing profile; the new lab is independent research dated 6 October 2026. Historical client-deployed quoting results are none established in the available inspected artifacts. Role-specific 20-hour commitment, sustainable four-hour UK overlap and interview slot remain to be confirmed before submission.
