Sarvesh Tiku

CachePilot: KV-cache-aware routing for LLM inference

Python, FastAPI, asyncio, SQLAlchemy, PostgreSQL, Prometheus, React, TypeScript, Docker

Routes requests to the worker that already holds their prompt prefix in cache, trading locality against load. In simulation it cut median time to first token by 81% and raised throughput 18% over round robin, and reproduced the failure mode where pure cache affinity overloads a single worker.

SwarmGuard: minimal-intervention containment for multi-agent AI systems

Python, Streamlit, sentence-transformers, Parquet, Docker  demo write-up

Reconstructs how an error or unsafe behavior may have spread between agents from their execution traces, as a time-respecting graph of agents, messages, and shared resources. It then scores candidate controls (blocking a channel, restricting writes, isolating an agent) by how much inferred spread they would have prevented against collateral disruption to benign activity. On a 20.5k-event wiki swarm, restricting writes to one relay page removed 97% of inferred paths at about 0% collateral. Built at the AI Swarm Dynamics Hackathon (AI Village × Grove Research).

Quorum: consensus-gated actions for AI agents

Python

Several models vote on an agent's proposed action before it runs, and anything without consensus is rolled back. Includes SDK integrations and failure injection tests to measure what the gate catches.

Pulse: a low-latency market data, execution, and risk engine

C++20, Linux

Ingests a binary UDP feed, recovers lost packets over TCP, rebuilds order books, and runs strategies through pre-trade risk into a simulated exchange, with deterministic record and replay. Order book updates take 10–13 ns with zero heap allocations; the threaded SPSC pipeline has p50 0.92 µs and p99 1.6 µs latency at 1M msgs/s. Threading did not raise throughput over the single-threaded version, which I think is the most useful thing it taught me.

Airlock: context integrity for coding agents

Node.js, Python and TypeScript SDKs, Linear API  demo

An agent can check its context, plan, and then act after a human has changed the rules. Airlock treats that as a time-of-check to time-of-use bug. Each agent gets a versioned lease on the facts it was given, and a deterministic gate re-checks freshness, consistency, provenance, authority, and conflicts with other agents right before any consequential action. Agents whose context goes stale are paused, and denied actions never run. An LLM reviewer can flag problems but can never override a denial.

Solar Surfers: ML solar siting and yield prediction for Colorado

Python, pvlib, scikit-learn, pandas, JavaScript  demo

Helps Colorado homeowners decide whether rooftop solar is worth it: for any address it estimates generation, forecast confidence, and 25-year financial value from 19 years of hourly weather at 150 sites. The day-ahead forecaster reaches 7.5% hourly RMSE (R² 0.94) on held-out 2024 and 7.7% at sites never seen in training. Built during the Computing for Good fellowship at Colorado School of Mines.

Flight Delay Prediction

Python, XGBoost, TensorFlow, Flask, React, Docker, Google Cloud Run

Delay prediction from 6M+ flights joined with weather data; compared XGBoost, random forests, MLPs, and RNNs.

Tracking Rare Earth Elements in Acid Mine Drainage

Python, C++ (pybind11), pandas, NumPy, ArcGIS, spatial interpolation

Hyperspectral imagery, field samples, and stream flow records combined to screen 25+ legacy mine sites for rare earth enrichment.