KM Kanav Mehta Quantitative Researcher

Agentic Research · Institutional Memory

The research ledger that refuses to let a lucky backtest ship.

Spark-MCP gives quant research teams a memory. It logs every hypothesis and outcome (failures included), deflates significance for the full trial count, and maps the unexplored corners of research space. Validated ideas graduate into replay-gated skills, with a human veto at every promotion.

Back to Projects GitHub

Trial discipline

Significance, deflated by construction.

  • Every trial is inserted before dispatch, so retries and reruns never double-count; the trial ledger is exact-once by design.
  • Every verdict pairs raw Sharpe with its Deflated Sharpe, adjusted for how many things the team actually tried.
  • Failures are first-class records: the ledger remembers dead ends so the next agent doesn't re-walk them.

Exploration

A map of what nobody has tried yet.

  • A MAP-Elites quality-diversity archive surfaces genuinely unexplored method combinations, not just nearest neighbors of past work.
  • Successful trajectories distill into challenger skills, promoted only after a replay gate and explicit human approval. The LLM narrates; it never decides.
  • A bounded artifact-evolution loop (hypothesis → backtest → DSR gate → computed promotion) stays bit-reproducible given seed, caps, and provider; held-out data never leaks into feedback.

In the stack

The judge beside the engine.

  • Spark never backtests or owns market data: the Alpha Engine computes (backtests, causal tests, risk decomposition) and Spark judges the result.
  • Engine-agnostic MCP peer: plugs into any MCP-capable agent, and stays fully usable offline in degraded mode with a lexical novelty proxy.