Parth Kshirsagar is a Software Engineer who builds full-stack web, mobile, and
AI/ML systems. 1st Place at the UC Berkeley AI Hackathon 2026 (The Token Company
track) for ReCompress, a query-aware context-compression model.
ReCompress
1st Place — The Token Company Track · UC Berkeley AI Hackathon 2026 · Jun 2026
A query-aware rewriting layer that extends compression into the regime deletion can't reach, distilled into a 1.5B model, then carried into multi-turn conversations to keep a 12-turn chat flat (184 tokens) while a naive agent balloons to 1,482.
- A 1.5B model that recovers the query-aware regime bear cedes, beating bear with statistical significance while emitting about 8.5x fewer tokens, and transferring to a near-in-distribution benchmark it never trained on (2Wiki, +46%; directional-but-unproven on the dissimilar OOD sets). It complements deletion rather than replacing it: deletion stays best for fast/verbatim/reusable; rewriting adds the query-specific case.
- We stress-tested our own headline before a judge could. The teacher and solver are both DeepSeek (a circularity a sharp reviewer attacks first), so we re-scored with an independent solver (Claude Sonnet): the gap is invariant, Delta vs bear = +0.288 (independent) vs +0.285 (in-family), CI excludes zero both ways. The result is not a same-family artifact.
- A unified system: the same distilled compressor works as a single-shot compressor AND as a multi-turn memory engine (the strongest backend of the three tested, vs DeepSeek and bear).
- Research-grade rigor plus intellectual honesty in 24 hours: bootstrap CIs on every claim, a cross-solver audit, a mask-the-answer audit measured against ourselves, 5 named experiments with a clear winner, three documented negative results, a conceptual finding (the "deletion ceiling"), a 13-figure visualization suite, and a custom multi-turn benchmark.
Stack: python, qwen2.5-1.5b, lora, unsloth, modal, h100, deepseek-api, the-token-company-sdk, huggingface-datasets, hotpotqa, 2wikimultihop, musique, squad, ms-marco, matplotlib
Links: Code · Demo · Devpost · Video · Paper · Slides · Colab · Paper DOI (alt)
CITED
Winner — Hexclave Track · c0mpiled-11 Startup School Hackathon · Jul 2026
Points an answer engine at your category, finds which sources it actually reads, tells you where you're missing — then opens the pull request that fixes it.
- Won the Hexclave track at c0mpiled-11 Startup School Hackathon (Transpose Platform, SF) by selling an AEO (Answer Engine Optimization) report to Hexclave
- Built Part A, the retrieval engine: runs category queries against answer engines, captures cited URLs, aggregates by domain, and scores visibility as cited_queries / total_queries
- Shipped an end-to-end pipeline that audits a site's AEO gaps, generates metadata/JSON-LD/robots/llms.txt fixes and opens a PR, plus a Voygr voice-call outreach loop that calls only non-competitor cited sources (LLM-classified)
Stack: TypeScript, Node.js, OpenAI API, Perplexity Sonar, Anthropic API, Octokit, Voygr, IndexNow
Links: Code · Demo
SwiperNoSwiping
Winner — Hexclave Track · c0mpiled-13 Hackathon · Jul 2026
A Chrome extension that hides the Buy Now button behind an invisible wall — to get through, you justify the purchase to an LLM that knows your budget, savings goal, and last twelve buys. Weak excuse → DENIED, and the money moves to your savings counter.
- Won c0mpiled-13 with a Chrome MV3 extension, built in 3 hours, that intercepts checkout buttons and makes you justify each purchase to a skeptical LLM before the click goes through
- Engineered a fail-open interrogation flow: a content script overlays buy/checkout/add-to-cart buttons, a FastAPI backend enriches product context with the user profile plus last 5 purchases from SQLite, and DeepSeek returns a verdict in at most 3 questions
- Denied purchases roll into a running savings total with a 10-minute per-item cooldown; the extension deliberately fails open so a checkout is never trapped when the backend or DeepSeek is unreachable
Stack: Chrome MV3, JavaScript, FastAPI, Python, DeepSeek, SQLite
Links: Code
AdLayer
3rd Place — Render · Zero Human Company Hackathon (Terac) · Aug 2026
The ad network for the answer layer — sells disclosed sponsored placements in publishers' llms.txt (the file AI agents read), enforces the [SPONSORED] label in code, then measures whether that disclosure survives when an answer engine repeats it.
- Won 3rd place (Render) at the Zero Human Company Hackathon by Terac (Humanmade, SF) — an ad network that sells disclosed placements in llms.txt, the inventory nobody had priced
- Enforced disclosure in code: every served block carries a [SPONSORED] tag via an assertDisclosed() guard that throws before any write, backed by a GLiGuard compliance veto and no opt-out flag
- Ran the novel measurement — poll answer engines after serving a labeled placement, then classify the outcome as surfaced_labeled, surfaced_unlabeled, or absent — reusing the citation-capture engine ported from CITED
Stack: TypeScript, Node.js, llms.txt, GLiGuard, Stripe, Render
Links: Code
Staple
In Development · Extension · Hackathon Project · Jun 2026
A browser extension where an animated pixel-art cat lives on any webpage, reads the DOM, and walks you to the exact UI element you ask for in plain language — built for anyone lost in confusing, ever-changing dashboards.
- Built a cross-browser extension (Chrome MV3 + Firefox MV2) where a cute pixel-art cat reads the page DOM and guides you to any UI element from a natural-language question
- Designed for accessibility — helps people struggling with confusing, hidden-menu dashboards find things (like where your OAuth token settings live) just by asking the cat
- Multi-provider LLM support (DeepSeek, OpenAI, Anthropic, Groq, Gemini, or local Ollama) maps user intent to a specific element; the cat scrolls to off-screen targets and explains in a speech bubble. Published extension coming soon
Stack: JavaScript, Chrome MV3, Firefox MV2, Multi-Provider LLM, DOM Parsing, Pixel Art
Links: Code
ThatsMyEcho
Best Domain Name (MLH/GoDaddy) · Bitcamp 2026 @ Univ. of Maryland · Apr 2026
A 2D platformer built in 36 hours where you play as a cat sound spirit. Movement costs sound energy. Your echoes bait enemies. Silence means death.
- Won Best Domain Name (MLH/GoDaddy) at Bitcamp 2026 building a Unity 2D platformer in 36 hours
- Built frame-by-frame echo recording system — up to 3 simultaneous movement ghosts that interact with the world and bait enemies
- Engineered custom platformer controller with coyote time, wall jumping, and pursuit AI that accelerates toward echoes rather than the player
Stack: Unity, C#, Rigidbody2D, Custom AI, Plastic SCM
Links: Code · Demo · Video
WarGames
3rd Place — $500 · Compiled-6 @ Penn State · Mar 2026
Fully autonomous multi-agent simulation. 5 AI agents, no human in the loop, exploring AI accountability through an extended Prisoner's Dilemma with permanent threat scores.
- Won 3rd place ($500) at Compiled-6 Hackathon hosted by Transpose Platform
- Engineered 5-agent system with permanent threat scores and asymmetric consequence weighting — no human in the loop
- Designed accountability tracing framework to identify which agent caused systemic breakdown
Stack: Vite.js, MCP, Multi-Agent AI, Game Theory
Links: Code
SonicMind AI
Recognition — Qualcomm Edge AI · Qualcomm Hackathon @ NYU Tandon · 2025
Real-time AI audio engineer built in 24 hours. Custom CNN achieving <10ms latency and 85%+ accuracy on Snapdragon X Elite. Transforms any space into a professional studio.
- Built real-time audio AI at Qualcomm Edge AI Hackathon (NYU Tandon) in 24 hours
- Fine-tuned 4 AI models and built CNN from scratch achieving <10ms latency, 85%+ accuracy on Snapdragon X Elite
- Engineered fully on-device processing with lower power consumption than streaming
Stack: TensorFlow, Snapdragon X Elite, Python, Custom CNN, DSP
Links: Code
404Found
Recognition — Hack Dearborn · Hack Dearborn 2025 · Oct 2025
99% automated privacy guardian. 5 specialized AI agents analyze bank statements, flag data brokers, and auto-generate GDPR/CCPA deletion requests.
- Built 99% automated privacy guardian using Fetch.AI Agentverse multi-agent system at Hack Dearborn 2025
- Orchestrated 5 specialized AI agents for OCR, classification, financial coaching, privacy risk, and request generation
- Merged cybersecurity and personal finance into a "last click" empowerment model with GDPR/CCPA compliance
Stack: Fetch.AI Agentverse, Gemini AI, MongoDB, Python, Tesseract OCR
Links: Code · Video
SNOM Robot
Hardware Category — Bitcamp 2025 · Bitcamp @ Univ. of Maryland · Apr 2025
Emotionally intelligent AI companion robot for neurodiverse children. Tracks moods through facial expressions and voice tone, adapts conversations in real time.
- Built emotionally intelligent robot at Bitcamp 2025 using Raspberry Pi 5 and OpenCV for autonomous person tracking
- Integrated Gemini 2.0 Flash for adaptive conversations helping neurodiverse children build social skills
- Engineered complete hardware-software system with real-time mood detection and parent dashboard
Stack: Raspberry Pi 5, OpenCV, Gemini 2.0 Flash, Flask, Python
Links: Code · Devpost · Video
Pathfinder AI
Recognition — HackPSU · HackPSU @ Penn State · Oct 2024
Dual-purpose AI platform with a career guidance chatbot and RAG-based document summarization. Zero hallucination through multi-step retrieval pipelines.
- Developed dual-purpose AI platform at HackPSU using Next.js and Langchain
- Built two distinct pipelines: curated career advice and RAG-based document summarization to prevent hallucination
- Deployed production-ready AI advisor praised by testers for reliability and accuracy
Stack: Next.js, React, TypeScript, Langchain, OpenAI, Vercel
Links: Code · Devpost · Video
VR Circuit Simulator
Best VR Hack · HackUMass @ UMass Amherst · Nov 2024
Immersive VR circuit simulator with realistic breadboards, NLP-powered circuit generator, and real-time physics simulation for electronics education.
- Won Best VR Hack at HackUMass with 500+ competitors
- Developed NLP-powered Circuit Generator for natural language command processing
- Built 2D Circuit Simulator in Pygame for rapid prototyping, integrated 3 tools into one ecosystem
Stack: Unity, Meta Quest, Python, Pygame, NLP
Links: Code
The Moth Challenge
1st Place — Cybersecurity · Bitcamp @ Univ. of Maryland · Apr 2024
Gamified cybersecurity puzzle platform inspired by Cicada 3301. 7 unique puzzles with progressive difficulty scaling for 550+ competitors.
- Won 1st place cybersecurity track at Bitcamp (University of Maryland, 550+ competitors)
- Developed 7 unique puzzles with modular Next.js + TypeScript architecture
- Implemented seamless Figma-to-React workflow with progressive difficulty scaling
Stack: Next.js, React, TypeScript, Figma
Links: Code · Devpost · Video
Likhit
Recognition — HackHarvard · HackHarvard @ Harvard University · Oct 2023
Handwriting improvement platform bridging paper and digital writing. AI-powered analysis using GPT-4 and OCR for adaptive learning and personalized phrase generation.
- Created AI-powered handwriting platform at HackHarvard 2023 (600+ participants)
- Built real-time p5.js canvas with Flask + OCR backend and GPT-4 adaptive learning
- Developed progress tracking with personalized feedback and data-driven insights
Stack: p5.js, Python, Flask, GPT-4, OCR
Links: Code
Zeph
YHack 2026 @ Yale · YHack 2026 at Yale University · 2026
Natural language network orchestrator for home lab machines. Runs Ollama locally, plans multi-machine workflows via LLM, dispatches commands in parallel. React dashboard + mobile PWA with voice input.
- Built local AI network orchestrator at YHack 2026 (Yale) — natural language commands dispatch to multiple Arch/Hyprland machines in parallel via LLM-planned workflows
- Engineered async multi-machine dispatcher with aiohttp, live WebSocket dashboard, SQLite device registry, and Flask agents with whitelist security
- Developed mobile PWA with tap-to-speak voice input and distributed notes system — fully local, zero cloud dependency
Stack: FastAPI, Ollama, Python, React, Vite, SQLite, WebSockets, Flask, PWA
Links: Code · Video