Skip to main content
ML & Software EngineerAkshat Vasisht
Back to Home

Projects

Selected Projects

Projects, hackathon demos, and open-source.

Sigil

Coding agents produce inconsistent results. Chaining them requires fragile scripts with no recovery.

A CLI that orchestrates coding agents as deterministic directed graphs. Validation gates enforce quality between steps, and checkpoints enable crash recovery.

Reproducible, resumable runs. Visual editor to compose and monitor agent workflows in real time.

Generative AIDevToolsSystems

Janus

Audio codecs bottom out around 6 kbps. Below that, low-bandwidth calls degrade into unusable noise.

A real-time codec that transmits speech as text and pitch data instead of waveforms, hitting 300 bps. Reconstructs natural audio preserving speaker identity via generative TTS.

20x below Opus minimum. Fully local inference, no cloud dependency. 158x satellite cost reduction.

Generative AIEdge AISystemsHackathon

CardinalCast

End-to-end ML class project covering real-time data ingestion, job scheduling, and automated payouts.

A weather wagering platform that prices bets using a calibrated probability model. A daily pipeline resolves wagers against NOAA actuals and refreshes odds.

Sub-100ms pricing. House edge held at 5-7% through calibrated probability models.

MLData EngineeringFull-Stack
PT

Prompt Tuner

Prompt optimizers either leak data to cloud APIs or ship static rules that go stale within months.

A Chrome extension that rewrites prompts across major LLM platforms entirely on-device. A quarterly pipeline scrapes official docs and refreshes the rule set automatically.

Fully offline, zero bytes leave the browser. Rules refresh automatically every quarter.

Generative AIEdge AIFull-Stack
PA

Prior Art

LLM-based package recommendations are non-deterministic, hallucinate, and vary per run.

An MCP server and CLI that evaluates packages across registries, scoring them on health signals and returning a consistent build-or-borrow verdict.

Covers PyPI, npm, crates.io, and pkg.go.dev. Algorithms draw from four peer-reviewed studies.

Data EngineeringDevTools
SS

SchedSec

AI schedulers generate conflicts: overlapping slots, broken dependencies, exceeded energy budgets.

A self-hosted scheduler on Cloudflare Workers. Resolves all constraints deterministically before handing validated slots to an LLM for natural-language placement.

100% constraint adherence by construction. Learns user preferences from schedule edits over time.

Generative AISystems

Trip Debugger

LLMs hallucinate tight connections and miss visa buffers. No travel planner offers a formal proof.

Uses LLMs to parse itineraries into formal constraints, then proves feasibility with an SMT solver. Maps failures to actionable fix suggestions.

Sub-200ms constraint solving. Interactive slider re-solves incrementally for real-time what-if tuning.

Generative AISystemsFull-StackHackathon

Contact

Reach Out

Happy to chat about opportunities, research, or interesting projects.

© 2026 Akshat Vasisht. All rights reserved.