Jesai Tarun

Co-Founder

Amherst, Massachusetts, United States12 yrs 5 mos experience

AI EnabledAI ML Practitioner

Key Highlights

Ranked among top 10 AI developers on GitHub.
Developed innovative on-device AI solutions.
Contributed to democratizing education through tutoring.

Stackforce AI infers this person is a SaaS-focused AI developer with a strong emphasis on educational technology.

Contact

Skills

Core Skills

Large Language Model Operations (llmops)Product ManagementMachine Learning

Other Skills

Python (Programming Language)OptimizationProgram PlanningScientific ExperimentationIB Computer Science Higher LevelIB Math AA Higher LevelIB Physics Higher LevelIB PsychologyFront-End DesignSocial SciencesTransformer ModelsComputer LiteracyComputer Science EducationDaVinci ResolveLaboratory Robotics

Experience

12 yrs 5 mos

Total Experience

5 yrs 6 mos

Average Tenure

3 yrs

Current Experience

Sentient os

Founder & Developer, Sentient OS (won $3000 grant; applying to YC S26 & a16z speedrun)

Sep 2025 – Present · 8 mos · United States

Sentient OS: a private, on-device AI layer for your entire digital life. Your phone and computer analyze every screenshot, file, bookmark, note, and email while they charge overnight; none of this data ever leaves your device. This unlocks three magical capabilities: talk to your data in natural language ("what was that wine I liked?"), ambient intelligence with proactive reminders (that tax return in Downloads before the deadline), and beautiful auto-generated knowledge graphs.
The bigger bet: via the Model Context Protocol (MCP), Sentient OS becomes the memory layer for every AI agent. Your ChatGPT and Claude finally know you; without any of that data touching the cloud. I'm building the memory infrastructure for the AI era.
Running a real Vision LLM on a phone is supposed to be impossible; these models are 8GB+ and iPhones give apps 4GB. Apple's own MLX framework doesn't support the smart compression needed. So I engineered my own k-quant-style selective quantization on MLX:
Multimodal projector (the model's "eyes") preserved at high precision; swapped in the eyes from a larger Qwen model for better image understanding
First and last embedding layers kept minimally compressed for coherent outputs
All remaining transformer blocks aggressively quantized
Fine-tuned on top of this custom base
Result: a 3.11GB Vision LLM on an iPhone, ~3 seconds per screenshot, entirely offline. My independent research paper on quantization effects on LLM reasoning (5,000+ runs; preparing for publication) directly informed this.
Why only on-device works: cloud competitors are structurally dead on arrival. Cloud vision APIs cost ~$15/user/month at this scale. Sentient OS costs $0.11 per user for 1,000 screenshots.
Working prototype: ~3,000 screenshots processed entirely offline. Closed beta in 3 weeks. Won UMass Tech Challenge ($3,000). Inbound press from AIM India. Active conversations with Entrepreneurs First ($250K SF founder track), a16z speedrun Alpha, and YC F26.

Large Language Model Operations (LLMOps)Product ManagementPython (Programming Language)Machine LearningOptimization

Independent research (scored top 5%); preparing for publication.

Research on Quantization Effects on LLM Reasoning

Mar 2025 – Present · 1 yr 2 mos

First comprehensive study on how quantization affects LLM reasoning. Created novel 654-question SAT evaluation dataset across 5 reasoning categories. 5,000+ systematic runs testing Llama 3.1 8B from 8-bit to 2-bit.
Found that reasoning degrades non-linearly with distinct breakdown thresholds; k-quants occasionally outperformed less-compressed traditional quantization. Built complete evaluation pipeline from scratch. Findings directly informed Sentient Screenshots' custom selective quantization!

Self-employed

Creator & Lead Developer, Writing Tools (featured in 28+ publications; was global top 10 AI dev)

Oct 2024 – Present · 1 yr 7 mos · United States

Ranked among global top 10 trending AI developers on GitHub in October 2024 -- GitHub's most competitive field with thousands of AI projects competing for recognition. During this time, Writing Tools was among the top 10 growing AI repos in terms of star count.
Developed an acclaimed system-wide AI writing assistant that outperforms Apple Intelligence and Grammarly, with real-time grammar correction, tone adjustment, multilingual support, and content summarization.
Featured in 28+ major worldwide tech publications and gained 2,000+ GitHub Stars. Empowering 30,000+ users.
Architected provider abstraction layer enabling seamless integration of any LLM API (OpenAI, Gemini, Ollama, other local inference engines…) with streaming responses.
Engineered advanced clipboard manipulation system to replace text across any application without affecting user clipboard; implemented multi-monitor screen-edge detection and intelligent popup positioning.
Led team of 8 contributors including senior engineers; orchestrated native macOS Swift port.
A user with ALS who types using only their eyes told me it changed their life.

Large Language Model Operations (LLMOps)Product ManagementPython (Programming Language)

Bliss ai

Creator & Developer, Bliss AI (Google Gemini API Developer Competition Nominee)

Jun 2024 – Present · 1 yr 11 mos

Created a novel Android app that uses fine-tuned Gemini to deliver hyper-personalized STEM tutoring in the local curriculum. Answered >20,000 questions for 100+ daily users.
Nominated in the Google Gemini API Developer Competition & secured ₹30,000 in API credits. Initial prototype won (3rd) in India's largest high-school hackathon, Oakridge Codefest 2024.
Deployed on resource-constrained school tablets at Indus International Community School and VIDYA Center of Excellence; extensively optimized for low-end hardware.

Cambridge centre for international research

Machine Learning Research with Dr. Hannah Rana, Harvard

Mar 2024 – May 2025 · 1 yr 2 mos · Remote

Developed a novel machine learning model using Bayesian Optimization to predict optimal cryocooler parameters from Oxford University experimental data.
Co-authored paper with Dr. Hannah Rana; aimed for peer-reviewed publication in the Journal of Cryogenics.

Scientific ExperimentationMachine Learning

Schoolhouse.world

Volunteer Tutor

Jun 2023 – Present · 2 yrs 11 mos · Remote

Having achieved a top 1% 1550 SAT score, contributed to democratizing SAT prep by offering free online tutoring to 100s of students from 29 countries.
Created and shared prep material, documents with tips, and personalised revision material for my students even outside of the tracked volunteer hours. Also helped Schoolhouse moderate the platform.

Program Planning

Self

Personal Projects & Technical Exploration

Jan 2014 – Present · 12 yrs 4 mos

Love pushing technology to its limits — featured in WIRED and 39+ other publications for getting iPadOS running on an iPhone (with a macOS-like experience when you plug in a monitor!)
First to investigate Apple's new neural accelerators in each GPU core.
Reverse-engineered macOS internals to get macOS working on a Dell XPS laptop. Open-sourced on GitHub (30+ stars).
First to extract system prompts from ChatGPT Advanced Voice Mode, Microsoft Copilot, Sarvam AI, GitHub Copilot, Dia Browser, etc.
Tinkering since age 8 — from jailbreaking iOS and compiling custom Android ROMs to getting Minecraft running on a smartwatch.