Software Engineer · UC Davis CE · Class of 2027

Saarth Kishore

I build production data platforms, LLM observability tools, and on-device AI — software that ships for real users, not slide decks.

9+
Campus data partners
3,600×
Pokémon ranking speedup
Live
Tone · Review AI · Aggies
Hands-free
Motion Sense Game

Builder who owns the pipeline

I’m a Computer Engineering student at UC Davis (Mathematics minor) who ships software end-to-end. At ASUCD Innovation & Research Lab I develop production systems for Aggies on the Move — unifying mobility feeds from 9+ campus partners into PostgreSQL warehouses and React dashboards used daily by researchers and staff.

I also ship AI products: Tone is an LLM reliability stack — production drift monitoring, per-conversation goal alignment, and llmtest (pytest for AI behavior) that turns a drifted chat into a CI suite — Review AI Detector is a Chrome extension + FastAPI stack that scores product and business reviews across 13 sites (listing trust, template clusters, posting bursts) — and Omnitrix is an always-on desktop assistant (screen-aware context, local Ollama, vector memory) engineered for privacy and speed.

On the games / CV side, Motion Sense Game is a shipped Unity 6 title controlled entirely by webcam hand gestures — MediaPipe landmarking, a dual-hand gesture state machine, and a WebSocket bridge into God of War–style third-person gameplay — no controller required — and Harvest Vault is a live Roblox farming simulator with pets, private estates, rebirth progression, and DataStore-backed saves.

Looking for teams that care about impact: internships, research labs, and ambitious product builds in full-stack, data platforms, LLM systems, computer vision, and interactive software.

Where I ship in production

Software Developer

ASUCD Innovation & Research Lab · Aggies on the Move

Live

Campus mobility data lived in siloed partner feeds. Researchers and transportation staff needed one trusted warehouse and dashboards — not another prototype.

  • Design warehouse schemas and migrations; ETL across 9+ partner formats with validation, retries, and alerting
  • Ship React dashboards for ridership, parking occupancy, and permit analytics used by non-technical staff
  • Tune PostgreSQL indexes and query patterns for researcher workloads; document ingestion playbooks for new partners
  • Coordinate field mappings and data-quality SLAs; debug production issues when feeds drift or APIs change
  • React
  • PostgreSQL
  • ETL
  • Data Pipelines
  • Analytics

Flagship products

Tone

LLM reliability stack — prevent, detect, diagnose drift

Live

Deployed LLMs don’t fail loudly — they drift. Tone is a three-layer reliability stack: catch behavioral regressions before merge, monitor production, and diagnose a single chat when it goes off-goal.

15
Runtime probes · 5 tone / 5 fact / 5 persona
≤5 min
Scheduled sample cadence
3
Drift signals · MMD + KL + cosine
7+
LLM presets · any OpenAI-compatible
3
llmtest assertion types
0.35–0.40
Category alert thresholds
  • Runtime monitor — 15 curated probes on a ≤5 min cycle; MiniLM + ChromaDB; MMD / KL / cosine heatmap alerts (category floors 0.35–0.40); detector unit tests separate identical vs shifted corpora
  • Chat review — paste a transcript or upload a ChatGPT export; per-turn goal-drift timeline, peak/first-alert, and steering tips from the connected LLM or Tone heuristics
  • llmtest — “pytest for AI behavior”: semantic equals / tone / excludes assertions, snapshot baselines, example 3-test suite, and a GitHub Action that fails PRs on drift
  • Generate llmtest — from a drifted chat, emit distinct probes with claim-level baselines (style personas + adversarial excludes), then copy-paste into CI — production signal → prevention
  • Connectors — presets for OpenAI, OpenRouter, Groq, Together, DeepSeek, Fireworks, Ollama, plus any OpenAI-compatible /v1 endpoint; JWT / Google auth; live at tonehq.xyz
  • FastAPI
  • React
  • TypeScript
  • ChromaDB
  • GitHub Actions
  • llmtest
  • Railway

Review AI Detector

Chrome extension + ML API for review trust on 13 sites

Live · Chrome Web Store

Shoppers can’t tell if reviews are human or AI-generated. Built a niche detector for short, informal product reviews — with a high-precision threshold so real customers aren’t falsely flagged.

13
Supported review sites
1.00 / 0.90
Precision / recall (AI) · held-out test
0.00
False-positive rate on humans
400
Reviews evaluated · 280 / 60 / 60 split
2
Chrome Web Store installs
0.72
Production AI threshold
  • Manifest V3 extension — per-site adapters across 13 shopping / review sites; listing trust card, per-review badges, hover signals
  • FastAPI on Render — calibrated sklearn hybrid + listing analysis: template clusters, posting bursts, adjusted ratings, visit history sparklines
  • Held-out hybrid test (n=60): precision 1.00, recall 0.90, FPR 0.00 at threshold 0.72; trained on 400 matched human/AI reviews
  • Production API at ai-comment-checker.onrender.com; live on the Chrome Web Store
  • Chrome Extension
  • FastAPI
  • scikit-learn
  • JavaScript
  • Python
  • Docker
  • Render

Omnitrix

On-device personal AI assistant

In development

Cloud assistants train on your screen data. Omnitrix keeps perception, reasoning, and memory on-device — always available, never uploading pixels.

  • Screen capture → OCR/CV → local Ollama LLM → Mem0 + ChromaDB → tray overlay
  • Active-window context, streaming responses, vector memory for habits across sessions
  • Global hotkeys and non-blocking overlay; offline-capable once models are pulled
  • Python
  • Ollama
  • OpenCV
  • OCR
  • ChromaDB
  • Mem0

Beyond Art

UE5 draw-to-manifest sandbox

Live on itch

Players sketch weapons, body parts, and tools that become playable gear. Sketch quality maps to stats — the better you draw, the better it works.

13
Forge categories
215+
Object classes
10–90s<1s
Score latency
1 PCPublic
itch.io Windows
  • In-game forge → PNG export → cloud Judge Bridge (validate + hybrid ink/AI score) → equip
  • Mode B: game on itch.io, judge on Render — gameplay and AI update independently
  • Melee, ranged, and body forges shipped; fairness floors/ceilings and anti-exploit guards
  • Unreal Engine 5
  • Python
  • Computer Vision
  • Render
  • Game Design

Motion Sense Game

Hands-free Unity action — webcam is the controller

Shipped

Controllers break immersion when you want to play standing up. Built a full dual-hand gesture pipeline so walk, combat, inventory, and a God of War–style camera all run from a laptop webcam.

  • MediaPipe Hands → static + dynamic classifiers → priority dispatcher → JSON WebSocket → Unity 6
  • Left hand: walk, jump, slash, punch, blast · Right hand: fist orbit look, inventory, pointer, accept/deny
  • Simultaneous two-hand input; grounded jump, attack feedback, and camera-relative locomotion shipped
  • Unity 6
  • MediaPipe
  • OpenCV
  • Python
  • WebSockets
  • C#
  • Computer Vision

Harvest Vault

Roblox farming simulator — crops, pets, private estate

Live on Roblox

Grow crops, hatch pets, and unlock a private estate. Full farm loop with rebirths, automation, world shops, and same-server estate visits — published on Roblox.

  • Server-authoritative plant / harvest, weather, mutations, and DataStore progression
  • Pets, gear tiers, auto systems, decor, and estate expansion up to 64 fields
  • Cohesive client UI — hotbar, shops, settings, SFX, and sell feedback
  • Roblox
  • Luau
  • DataStores
  • Gameplay Systems
  • Economy Design

Other builds

Spotify Clone

Full-stack streaming client

Spotify-style client on the official Web API — playback, queues, and browser-side voice ID via FFT.

  • OAuth 2.0 with silent refresh; search, scrubber, queue, session persistence
  • Microphone → Web Audio FFT → fingerprint match entirely in the browser
  • JavaScript
  • Web Audio
  • OAuth
  • Spotify API

Pokémon GO Analyst

ML + computer vision rankings

Competitive ranking used to take ~3 hours of spreadsheets per league. Automated to ~3 seconds (3,600×).

  • Video ingest → OpenCV frames → meta scoring for Great / Ultra / Master
  • Top-500 ranked lists + CSV export; reproducible runs
  • Python
  • OpenCV
  • ML
  • Video

GDSC Davis Housing

Campus housing discovery

Students compared apartments via scattered posts and outdated sheets. Built one trusted discovery app with GDSC.

  • React + Firebase Auth/Firestore: reviews, ratings, filters by price and amenities
  • Iterated with 50+ student testers before launch
  • React
  • Firebase
  • Auth

Mini-Golf Engine

From-scratch game physics

No middleware physics — every bounce and ramp is hand-tuned collision math at a locked 60 FPS.

  • Circle–segment collision, friction ramps, fixed timestep, config-driven levels
  • Shows performance-critical loop discipline beyond CRUD apps
  • Python
  • Pygame
  • Physics

Technical toolkit

Languages

  • C
  • C++
  • Python
  • Java
  • JavaScript
  • TypeScript
  • Node.js
  • Next.js

Frontend

  • React
  • HTML
  • CSS
  • Dashboards
  • Rapid Prototyping
  • Web Audio

Backend & APIs

  • REST APIs
  • PostgreSQL
  • SQL
  • ETL
  • Pytest
  • System Design
  • MongoDB
  • CI/CD
  • GitHub Actions
  • Git

Data & AI

  • Machine Learning
  • Computer Vision
  • OCR
  • LLMs
  • Embeddings
  • RAG
  • ChromaDB
  • Ollama
  • llmtest
  • Chrome Extensions
  • OpenCV
  • NumPy
  • Pandas

Game Development

  • Unreal Engine 5
  • Unity 6
  • Roblox / Luau
  • Sandbox Design
  • Gameplay Systems
  • Physics
  • Procedural Generation
  • Pygame
  • Blueprints
  • MediaPipe
  • Gesture UX

UC Davis

B.S. Computer Engineering · Minor in Mathematics

Expected graduation June 2027 · Class of 2027

Systems, algorithms, architecture, and hardware — paired with production software at ASUCD IRL and shipped personal products in data, ML, and games.

Core coursework

  • DSA, Algorithms, OOP
  • Computer Architecture & Networks
  • Circuits I & II
  • PDEs & Abstract Math

Let’s talk

Open to software engineering internships, research collaborations, and ambitious product builds. Based in Davis, CA — willing to relocate.