# Loukik Naik

**Software Engineer** — San Francisco, CA, United States

> I’m a software engineer who enjoys building systems that ship and scale. I’ve worked across backend, systems, and ML-adjacent infrastructure, building APIs, real-time services, and production pipelines. I like owning problems end-to-end and keeping systems simple and reliable. I’m especially interested in: - Backend & systems engineering - ML / AI-powered products - Distributed systems & infrastructure MS in CS from UC San Diego. Always learning, building, and happy to connect.

Career Archetype: **Junction Artisan** (HE-MX) — One of the first Junction Artisans on Saywise

## Links

- LinkedIn: https://linkedin.com/in/loukiknaik

## Experience

### Software Engineer, Eudia (2025-08 – 2026-03)
Palo Alto, California, United States · • Owned the Mergers & Acquisitions backend service end-to-end, building and operating the core infrastructure for an agent-based AI system for large-scale legal document analysis in production. • Partnered closely with lawyers and product managers to design and ship product features, translating legal workflows and requirements into scalable platform capabilities. • Designed and shipped backend APIs serving as the control plane for AI agent execution, managing document ingestion, run orchestration, result persistence, and retrieval of structured and tabular outputs. • Re-architected agent orchestration by migrating from Airflow to Temporal, enabling reliable execution of long-running, stateful agents with retries, idempotency, and deterministic behavior. • Built backend capabilities for multiple AI agents, including flexible document analysis, configurable PII redaction, Contract Comparison, and tabular review with schema-based inference and cell-level operations, backed by structured logging, workflow tracing, telemetry, and Kubernetes-native monitoring. • Built workflows coordinating multiple services, queues, and workers, addressing production challenges including duplicate execution, inconsistent state across retries, and downstream dependency failures. • Designed and managed Postgres database schemas and migrations using Alembic, delivering backward-compatible schema changes to support new agent capabilities. • Converted the service to a multi-tenant architecture, enforcing tenant-aware routing, strict isolation, and safe rollouts across live customers. • Managed the rollout of new backend features across staging, internal, and production environments using Kubernetes, Helm, and ArgoCD, supporting multiple customer deployments.

### Contributor, OpenFilter (2025-05 – 2025-08)
United States · OpenFilter is a universal abstraction for building and running computer vision workloads

### Machine Learning Engineer, Plainsight Technologies (2025-01 – 2025-08)
San Francisco, California, United States · • Developed and deployed real-time image segmentation APIs using FastAPI for vision models including SAM and MobileSAM, enabling seamless integration into internal applications. • Containerized services with Docker and automated infrastructure deployment via Terraform and CI/CD pipelines, ensuring scalable, consistent rollouts. • Designed a model caching system using symlinks and config-driven scripts, reducing friction during builds and cutting cloud runtimes by avoiding redundant model downloads at runtime. • Created a Cron-triggered retraining pipeline that detects newly labeled data and launches Vertex AI training jobs, enabling continuous model updates. • Designed an OCR model evaluation framework utilizing edit distance, confidence scoring, and text similarity metrics to automate quality checks pre-deployment. • Designed and implemented a robust integration testing framework to validate end-to-end performance of the model generation pipelines for segmentation, object and keypoint detection tasks.

### Machine Learning Intern - Computer Vision, Plainsight Technologies (2024-06 – 2024-12)
San Diego, California, United States · • Implemented serverless functions on Google Cloud Functions for real-time video processing. • Developed and orchestrated Kubeflow pipelines on GKE for video preprocessing and filter evaluation, triggered by GCS data ingestion. • Created an environment calibration model for deriving environmental statistics from video streams. • Automated testing and deployment using GitHub Actions, PyPi, and Docker for production readiness. • Ensured reliability through end-to-end tests and continuous integration.

### Graduate Research Assistant, UC San Diego (2024-05 – 2024-06)

### AI Developer, Lab Systems (I) Pvt. Ltd. (2023-02 – 2023-05)
Mumbai, Maharashtra, India · • Developed video forensics tool mvp for law enforcement agencies to analyse CCTV footage efficiently. • Researched different techniques of object and text recognition to improve performance of the software. • Leveraged Python libraries like pytesseract, easyocr, yolo, mediapipe, etc for object and text recognition

### Software Engineer, Hexaview Technologies Inc. (2022-07 – 2022-08)
Mumbai, Maharashtra, India

## Education

### Master of Science - MS in Computer Science, UC San Diego (2023-09 – 2024-12)
Specialization: Machine Learning Relevant Coursework: Fall 2023: CSE 202 - Design and Analysis Of Algorithms CSE 250A - Probability and Reasoning CSE 258 - Recommender Systems and Web Mining Winter 2024: CSE210 - Software Engineering CSE251U - Unsupervised Learning CSE251A - ML: Learning Algorithms Spring 2024: CSE256 - Natural Language Processing CSE224 - Graduate Network System COGS209 - Data Analytics/Stats Learning Fall 2024: CSE291 - Advanced Data Driven Text Mining

### Bachelor's degree in Computer Engineering, University of Mumbai (2019-10 – 2023-05)
Department Rank 1 - SLRTCE

### Pace Junior Science College (2017-01 – 2019-01)

## Projects

### Verdict - Synthetic Market Research Agent
- Built a synthetic focus-group engine that runs 200–2,000 Claude personas against a product, returns a 1 - 5 purchase-intent PMF via structured output, clusters qualitative themes, and drops an Opus 4.7 executive brief - FastAPI, Sonnet 4.6, SQLModel, React + TypeScript end-to-end. - Designed a Temporal-shaped async workflow with ContextVar-based token attribution, so concurrent simulations in one process correctly bill each Claude call back to the right sim_id without threading IDs through every callsite. - Replaced an embedder-based SSR scorer (which clustered everything into a 3.0–3.6 band regardless of sentiment) with an LLM scoring path: Sonnet 4.6 reads each multi-turn interview, returns the full intent distribution, and grounds personas in live market data via the web_search tool. - Shipped reusable persona libraries, A/B variant comparison with side-by-side KPI deltas, per-call cost/latency telemetry (~$1 per 10-profile sim), and a Vite + React UI auto-deployed to verdict.loukik.dev — inspired by the PyMC × Colgate 2025 LLMs Reproduce Purchase Intent paper.

### Podclipper - Long Podcasts to Reels in a Click
• Built a local pipeline that turns long-form podcasts into vertical 9:16 reels — Whisper, Claude, YOLOv8, and OpenCV chained end-to-end. • Designed a stage-based architecture with content-hash-keyed caching, so iterating on crop or prompt logic runs in seconds instead of re-transcribing. • Solved active-speaker tracking across hard shot cuts via per-shot x-center clustering with look-ahead crop seeding and mouth-motion → diarized-turn linking. • Shipped an LLM-as-judge quality gate, a React + Vite landing page, and GitHub Actions auto-deploy to podclipper.loukik.dev.

### DraftIn - Keyboard-Driven LinkedIn Outreach Drafter
- Built a local-first Chrome MV3 + Node pipeline that drafts personalized LinkedIn outreach from a screenshot buffer - Alt+K stacks viewports across LinkedIn profiles and JD pages, Alt+L ships them through Oracle (browser-driven ChatGPT, no API key, no SaaS), and the result lands on the clipboard via pbcopy. - Designed it screenshot-as-source-of-truth - no DOM parsing, no LLM tokens in the browser — and used activeTab + captureVisibleTab so the same buffer works on any page (career sites, founder pages, anywhere the user scrolls). - Solved Chrome MV3's clipboard-after-focus-loss limitation by writing through the local server with pbcopy / xclip / clip instead of fighting the service-worker clipboard restriction from a toolbar gesture. - Tuned Oracle browser-mode flags (--browser-min-stable-ms 15s, 10-min ceiling, 4-min recheck) so mid-stream pauses during image analysis and company web-search lookups don't trip premature capture; shipped a React + Vite landing page with GitHub Actions auto-deploy to draftin.loukik.dev.

### 3D Vehicle Visualizer (Digital Twin + API)
• Built a browser-based 3D vehicle viewer with a FastAPI backend that controls vehicle state (speed, brake lights, doors) and syncs it with a React + Three.js frontend. • Designed a digital-twin architecture where the backend owns state + configuration, keeping the UI purely as a renderer. • Solved real-world GLB inconsistencies via semantic part mapping, custom hinge/axis overrides, and per-model configuration stored via API. • Implemented production-style UX patterns: optimistic updates with rollback, debounced writes, and polling for multi-client consistency.

### Synapse – Decision-Based Spaced Repetition (iOS)
• Built an offline-first iOS app that trains engineering intuition using scenario-based flashcards focused on tradeoffs instead of memorization. • Implemented a modified SM-2 spaced repetition engine with confidence-based scheduling and per-scenario difficulty tracking. • Designed a JSON-based import pipeline to convert LLM-generated content into structured study decks with validation and mapping to SwiftData models. • Architected clean separation between DTOs and persistence models to handle complex object graphs and ensure scalable content ingestion.

### AutoReel - Multi Agent Reel Generator
• Built AutoReel, a side project that turns a single prompt into a ready-to-share 1-minute reel. • Set up a multi-agent pipeline (LangGraph): one agent writes the script, another generates images, another adds TTS audio, and the last stitches everything into video. • Used OpenAI models for script, voice, and visuals, and MoviePy for editing. • Designed the workflow to run end-to-end automatically, from idea → script → video — with no manual editing.

### Prompt Based Email Classifier
• Built a smart Gmail organizer that uses a custom prompt via LangChain to classify emails through either Ollama (local models) or the Gemini API (cloud) for real-time classification into categories like Work or Finance giving users full control over privacy vs. accuracy. • Designed a fast, responsive system using multithreaded Python (Flask) and the Gmail API to fetch and batch-process emails, with a React frontend for live updates and intuitive filtering.

### Surfstore: A Distributed File Storage System
• Built a Horizontally Scalable File Storage System in Go, storing data across multiple block and metadata servers. • This project makes use of a .proto file to generate grpc code so that the client and server stub can use different programming languages and can belong to different devices. • Implemented Raft consensus algorithm to sync changes across multiple metadata servers.

### PPG Signal Based Hypertension Prediction
• Developed robust ETL pipeline to process the MIMIC-III dataset ensuring high data quality and integrity for downstream analytics. • Achieved a test accuracy of 74% and a sensitivity of 91% on the MIMIC-III dataset.

### TritonHTTP: A Server written in Go
• Implemented a Go-based HTTP server from scratch to handle multiple virtual hosts, each serving content from distinct document roots. • Leveraged Go's concurrency model to handle multiple client connections simultaneously using goroutines. • The server parses the incoming requests and, based on the specifications of TritonHTTP, generates and serves 200, 400, and 404 responses.

### Active Amigo - Exercise Buddy App
• Our team developed an Android app that helps users find others with similar schedules for working out or participating in physical activities. • The app uses Firebase Firestore as the backend and includes an end-to-end CI/CD setup with GitHub Actions.

### Performance Analysis of Co-ordinate Descent vs Logistic Regression
• In this analysis study I performed a brief comparison of the performance of Logistic regression and 3 different varieties of co-ordinate descent i.e. Random Co-ordinate Descent, Iterative Co-ordinate Descent and Gradient based Co-ordinate Descent on the wine data set from sk-learn. • Although, the co-ordinate descent failed to achieve log-loss values comparable to logistic regression, they converged faster and had reasonable loss and accuracy values to be considered for practical use.

### Prototype Selection Technique For 1-Nearest Neighbour
• Compared and proposed different methods for prototype selection on huge datasets to summarize the data such that 1-NN model yields optimal test results. • Achieved identical test accuracy of 96.7% using only about 11% of MNIST dataset after prototyping.

### Anime Recommendation Website
Developed a website that suggests similar animes to a given anime entered by the user. Implemented a cosine similarity algorithm to calculate the similarity between anime vectors.

## Skills

Agentic Workflows, STT, Whisper, FastAPI, Swift (Programming Language), iOS Development, Mobile Application Development, Xcode, 3D Modeling, Video Processing, Software Infrastructure, Multi-agent Systems, Airflow, Temporal, LangGraph, TTS, Text-to-Image Generation, LangChain, LLaMA, Flask, Prompt Engineering, Multithreading, Routing Protocols, Large Language Models (LLM), Google Kubernetes Engine (GKE), Cloud Infrastructure, Kubeflow, Google Cloud Platform (GCP), Docker, Computer Vision, Machine Learning, Data Visualization, Remote Procedure Call (RPC), Concurrent Programming, Computer Networking, Go, Data Synchronization, HTTP, Problem Solving, Programming, Computer Science, Data Structures, Algorithms, Natural Language Processing (NLP), Software Project Management, Data Preparation, Data Analytics, Data Science, Statistics, Model Evaluation

## Credentials

- **SnowConvert for Developers** — Snowflake (issued 2026-01)
- **Generative AI: Working with Large Language Models** — LinkedIn (issued 2025-09)
- **Develop a Synthetic Monitoring Platform** — Build Fellowship by Open Avenues (issued 2024-09)
- **Software Engineer Intern** — HackerRank (issued 2024-03)
- **Neural Networks and Deep Learning** — Coursera (issued 2022-01)
- **SQL(Intermediate)** — HackerRank (issued 2022-01)
- **Codechef Snackdown** — CodeChef (issued 2021-12)
- **18th Rank | Code Wars** — Coding Ninjas (issued 2021-03)
- **Data Structures in Python** — Coursera (issued 2020-06)

---
Source: https://saywise.com/member4037 (last modified 2026-10-06T02:57:12.279Z)