Lochan Reddy

01. About Me

I'm Lochan, an undergrad student at Berkeley. I care more about the people I work with than the products I build.

I believe that having high agency is what separates the good from the great. It's a mantra I live by.

I build things that are actually useful, often stemming from personal issues I have faced. I love when my builds are used by real people and they find value out of it.

Languages

Python
Swift
C / C++
Java
JavaScript
SQL
Bash, R, Verilog

Frameworks / Libraries

PyTorch & PyTorch Geometric
TensorFlow
Scikit-learn
SwiftUI / React / Flask
Pandas / NumPy / NLTK
OpenCV

AI & ML

Diffusion Models
Fine-tuning (LoRA)
On-Device LLMs (Apple Intelligence)
GNNs
Temporal Modeling
Computer Vision

Tools / Platforms

Git / Docker
AWS SageMaker
Xcode
Tableau
Postman
Linux (Ubuntu) / Yosys

Currently studying at UC Berkeley

CS + DS Double Major

3.75 GPA

02. Experience

Software Development Engineer Intern

Amazon — Alexa Team

May 2026 – Aug 2026

Built modular C++ engines for real-time inference (wakeword verification) on the Alexa team — 64% latency reduction over the legacy system.

C++ Real-Time Inference Alexa AWS

Machine Learning Research Intern

Outsampler — Financial AI Research

Dec 2025 – May 2026

Fine-tuned a 3B-parameter multimodal SLM (Qwen2.5-VL) with LoRA to distill reasoning from GPT-4o, reaching 94%+ citation correctness for safety-critical model interpretability in high-stakes financial contexts.
Designed hybrid objective functions (ℒ_ITC) aligning spatio-temporal embeddings with analyst reports; improved OOD detection across 7 anomaly families (trend, frequency, noise, and others) by 15%.
Benchmarked against Unsupervised Random Forest via the ADBench protocol, evaluating AUC-ROC across 26 real-world datasets to validate deployment readiness and generalization across low-resource environments.

PyTorch LoRA Qwen2.5-VL Multimodal SLMs OOD Detection

ML Engineering Intern

Movicorn — Media & Entertainment

Jan 2026 – May 2026

Built an automated product placement pipeline using SAM 2 for instance segmentation and CLIP for zero-shot asset selection, eliminating manual frame-by-frame editing and enabling scalable branded content production.
Engineered a spatio-temporal warping engine via Farneback optical flow and Delaunay triangulation to maintain 3D geometric consistency; enforced motion coherence with FFmpeg/PyAV and deployed as a Flask REST service.
Architected the full inference pipeline from asset detection to composited frame output, supporting real-time branded content generation across arbitrary video streams.

SAM 2 CLIP Computer Vision Flask FFmpeg

03. Projects

Nanoserve — LLM Inference Engine

nanoserve

Zero-dependency C++ inference engine and serving system for Qwen2.5-0.5B, written from scratch: greedy decoding bit-identical to HuggingFace, paged KV cache (21x lower memory), and an OpenAI-compatible SSE server with continuous batching (2.85x aggregate throughput).
CPU decode optimized 14x via NEON SIMD, threading, and allocation-free decode, then 2x further with INT8 quantization (integer-sdot GEMV kernels, verified against fp32 goldens) — within 8% of llama.cpp on a reproducible benchmark harness.
Metal GPU backend with zero-copy unified-memory weight residency; documents an honest negative result — dispatch latency dominates GEMV at 0.5B scale, so CPU stays the default.

C++20 NEON SIMD Metal INT8 Quantization pybind11

IntentGate — iOS Productivity App

Feb 2026 – Present

Personal iOS app intercepting phone usage at the session level — before opening any guarded app, the user must state their intention; all apps unlock simultaneously for that duration, then hard re-lock.
Modular multi-target system (main app + 3 system extensions) communicating across sandbox boundaries via shared App Group; integrates Apple Intelligence (FoundationModels) for on-device intent classification with fallback to OpenAI/Anthropic/Gemini APIs or manual mode.
Structured output schema using @Generable forces the on-device LLM into typed fields (category, duration, rationale); every session persisted with SwiftData to surface accountability data in a history view.

Swift SwiftUI FoundationModels SwiftData FamilyControls

AI Product Placement Pipeline

aibranding

Automated pipeline that inserts brand assets into video scenes without manual frame-by-frame editing; uses SAM 2 for precise instance segmentation and CLIP for zero-shot asset selection from a brand library — no manual asset tagging required.
Tracks objects across frames with Farneback optical flow; warps logo geometry via Delaunay triangulation to maintain 3D geometric consistency; enforces temporal coherence through FFmpeg/PyAV.
Served as a Flask REST API with Docker for integration into production branded content workflows.

SAM 2 CLIP OpenCV Flask Docker

DistillAnomaly

Time Series Anomaly Detection

End-to-end Slurm-based research pipeline for time series anomaly detection on HPC (H100 GPUs); fine-tunes Qwen2.5-VL with LoRA using GPT-4o-distilled annotations across 7 synthetic anomaly families (trend, frequency, noise, and others).
Evaluates AUC-ROC against Unsupervised Random Forest baselines via the ADBench protocol across 26 real-world UCR datasets; aggregates citation correctness and metric summaries for deployment readiness validation.

PyTorch Qwen2.5-VL LoRA Slurm / HPC

Dark Pattern Auditor

Apr 2026 — Perplexity Hackathon

Full-stack agentic legal audit tool that autonomously navigates signup, cancellation, and cookie-consent flows using Playwright browser automation, capturing screenshots and page text at each step.
Classified detected patterns against live FTC and EU statute text using Perplexity sonar-reasoning-pro with real-time fetch_url tool calls; surfaced real enforcement precedents with case names, outcomes, and source links.
Streamed live crawler progress via FastAPI SSE to a Next.js frontend rendering structured legal audit reports with severity badges, screenshot thumbnails, and statute excerpts.

FastAPI Playwright Perplexity AI Next.js TypeScript Tailwind CSS

Finder+

finder-plus

macOS Electron app that overlays Finder with a smarter file browser and Claude-powered natural language search — "Find my tax documents from last year" just works.
Auto-indexes Downloads, Desktop, and Documents; supports real-time filename/tag filtering, file pinning, custom labels, and a global hotkey (Cmd+Shift+Space) to toggle the overlay from anywhere on macOS.

Electron React JavaScript Claude API

Financial Planning AI

Sep 2025 – Nov 2025

Full-stack investment platform combining FinBERT sentiment analysis with a GNN (PyTorch Geometric) modeling ETF/sector correlations as a graph; trained and deployed on AWS SageMaker.
Backtested 5 years (2018–2023): Sharpe ratio of 1.1, 10–12% return improvement during volatile periods, 82% Monte Carlo success rate for retirement goals — 15% above standard 60/40 portfolios.
~50ms end-to-end inference via model caching and async fetching; React + MySQL frontend for real-time portfolio rebalancing and personalized plan generation.

FinBERT PyTorch Geometric AWS SageMaker React MySQL