Alexey Zaytsev

Some Client projects: • AI Storyboarding Platform (Python backend): Integrated new image generation providers and refactored the generation abstraction layer across use cases (assets, scenes, inpainting). • Earnings Call Analysis & CEO Training: Built a transcription pipeline (AWS Transcribe / Deepgram) with speaker diarization and LLM-driven financial metrics extraction. Companion CEO training module generates realistic analyst questions via TTS from call scripts, transcribes the CEO's spoken responses, and provides feedback on both substance and delivery (WPM, filler words). Also built and maintain several open-source AI applications: • Congusto Chat: Multi-provider multi-model AI chat UI offering ultimate flexibility. SvelteKit, Drizzle/PostgreSQL, Vercel AI SDK, Lucia auth, S3/CloudFront for media. • lovely-docs: Documentation dehydrator for AI agents. Scrapes and clones codebases to generate recursive LLM doc summaries. SvelteKit UI, `npx` CLI, MCP server. • Pelican: LLM-driven SVG/ASCII art gallery with iterative refinement (resvg + Jinja2 prompts). Client-side inference (Vercel AI SDK), SvelteKit, PostgreSQL, S3/CloudFront. • latent-tools: PyTorch ComfyUI nodes for mathematical manipulation of diffusion latents (tensor blending, concatenation, noise generation) in SD/SVD. • svelte-asciiart: Svelte 5 library rendering ASCII art as scalable SVGs. Computes precise viewBox constraints and bakes CSS styles for standalone exports.

Led development of the generative AI platform supporting image, video, and 3D model generation. Served tens of thousands of users, generating hundreds of thousands of images: • Architecture & Inference: Built an end-to-end generation pipeline. The frontend talked to the backend that talked to Supabase/PostgreSQL, which queued tasks via RabbitMQ. Distributed GPU workers picked up the jobs, generated the media, uploaded to S3 (served via CloudFront+imgproxy), and pinged a backend webhook, which finally streamed updates to the client via Server-Sent Events. • Model Fine-tuning: Extensive work with Stable Diffusion inference, custom model fine-tuning, and LoRAs to optimize generation quality and capabilities. • Custom NSFW Detector: Off-the-shelf filters were unusable in production due to false-positive rates. Trained a custom EfficientNet classifier on a large, intentionally generated dataset, achieving production-grade precision. Integrated directly into the GPU worker pipeline. • MLOps & DevOps: Orchestrated and scaled the distributed inference infrastructure.

Independent contractor focused on AI apps and prototypes. Some that come to mind: • Dog Finder: An experimental missing-dog app exploring both facial recognition and noseprint uniqueness. Trained EfficientNet with triplet loss. • SAT Practice Generator: A dynamically updating test prep system seeded by real data. Built with Streamlit, Airtable, and OpenAI LLMs—migrating from GPT-3.5 to GPT-4 largely solved our initial correctness and consistency struggles. • GPTEditor: An AI-assisted copywriting tool—like Cursor, but for prose. Users highlight a paragraph and instruct the LLM on how to rewrite it. Built with SvelteKit, Tiptap, and Supabase.

Deep learning and AI intensive study period focused on understanding the fundamentals: • lovely-tensors: Built an open-source PyTorch debugging library for human consumption, which saw wide adoption among AI researchers. Later expanded to lovely-numpy and lovely-jax. • GPT-2 from Scratch: Built and trained a GPT-2 style transformer using raw PyTorch. • Autograd on NumPy: Implemented a custom autograd library on top of NumPy, featuring enough operators to train a miniature GPT-2 variant on the TinyStories dataset.

Smart solar grid tech: designed the main grid hub (Solar → Battery → regulated DC loads and high voltage sharing), a debugging adapter, and auxiliary internal boards. Schematics, PCB design, DC-DC converters and analog monitoring circuits. Led a small hardware team (3 engineers).

Dozens of client projects across digital, analog, and power electronics — from initial concept to production-ready designs and manufacturing oversight. Clients in medical devices, scientific instruments, food and consumer electronics. Designed and set up end-of-line automated testing, trained factory workers, and oversaw manufacturing in Shenzhen factories.

EdTech startup providing affordable laboratory equipment for molecular biology education in schools. Designed spectrophotometer, electrophoresis and transilluminator — hardware, electronics, mechanical enclosures, and embedded software. Co-founded with fellow students from the CRI Paris program.

Developed the first iteration of a low-cost, open-source spectrophotometer — the project that later evolved into the Koi.science startup.

Characterizing bacteriophage origin of replication region, exploring the potential of applying genetically engineered λ-phages to antibiotic-resistant bacterial infections in burn victims. Contributed to a published paper.

Researching the influence of electric fields on mammalian cell mobility and wound healing.

Virtio drivers for network and block devices, general kernel maintenance, bugfixes and development on illumos (OpenSolaris fork).

Linux kernel development for high-bandwidth telecom equipment. AltEll: 10 Gbps network routers. Protei: telephony over dedicated lines, WiMAX and GPRS/UMTS base stations. Device drivers, bootloaders, and board support packages on MIPS64, ARM and x86.