Hacker
Atharva Maskar
@atharvamaskar_8548
Elsewhere
Projects
Deflate : Grounded AI Swarms for Quant Teams.
Flip the Switch HackathonNov 2026
Alpha decay in quant algorithms used to take months. Now it takes days. Converting a vague market hypothesis into code is nearly free, so any edge gets copied and arbitraged away fast. Generic AI swarms make it worse: every extra backtest raises the best Sharpe you can get by pure luck (López de Prado, AFML §11.3). So a leaderboard of machine-generated strategies mostly ranks noise. My last swarm produced 259 candidates and promoted 0, and it took me months to see why. Deflate is a swarm of research agent built specifically for quant teams that keeps the speed of a swarm and refuses its false discoveries powered by Switch - agents read papers, propose hypotheses, write specs and explain results. They never decide whether a strategy is real. A deterministic gate decides: 16 checks from AFML, including purged CPCV, deflated Sharpe with the family's real trial count, PBO, an exposure-matched null, costs, concentration and drawdown. Agents cannot edit the gate. Every run is a counted trial, a failed idea gets at most four one-rule patches, and a fork inherits its family's trial count, so the bar never resets. A strategy that passes gets one holdout touch, then a named human signs off in Slack. Approved strategies go to paper trading, where CUSUM and three other decay monitors retire them the moment the edge fades. It's for quant researchers and small trading teams who already use AI agents and need to know which ideas are real before they reach live markets. In today's run the agents tested 8 strategies and the gate rejected all 8. That's the point: I'd rather get the reality check in testing than in live markets.
- Top 50
MochiAI
Graph HacksOct 2026
Support agents at a stock broking platform start every ticket with no context. The customer's history sits in old tickets, the fix that worked last week sits in a colleague's head, and help articles promise turnaround times the team stopped meeting months ago. Most tickets cross several teams, such as onboarding, KYC and the KRA, so the context an agent needs is spread across people and records that are never linked. AI assistants forget between sessions, miss how customers, issues, fixes and teams relate, and can't show the reasoning behind an answer. Mochi keeps one shared company memory in a FalkorDB graph. It briefs the agent before they start, shows the path behind every claim, and keeps itself current overnight.