Projects

Everything I've shipped, most recent first.

How it unfolded at Sandwatch
early 2025
The C# engine

Sole developer of a node-based engine: topic in, finished long-form video out, no human in the loop.

2025
Delivery & infra

A Discord bot drops the videos into channels; Terraform stands the whole platform up in prod and dev.

late 2025 → 2026
The creator platform

Thirty-plus REST endpoints, then the money: atomic credits, card + crypto checkout, exact-source refunds.

2026
Voices

New TTS provider, a cloning flow, and a catalog that grows from a handful of voices to 600 across 12 languages.

2026
Hardening

Multi-provider fallbacks halve the failure rate; GPU workers handle the audio and video post-processing.

2026
The rebuild

Generation extracted into a durable Inngest pipeline, graded by LLM judges, shipped with zero-downtime blue-green deploys.

All projects

AI Video Generation Pipeline

Primary engineer on the video-generation service — a dedicated service we extracted the generation logic into and rebuilt on Inngest. It's a ~20-step durable pipeline spanning script, voice, character, and multi-stage visual generation; each step suspends until its provider webhook fires, so a job can wait on a render as long as it needs and resume exactly where it left off without tying up a worker or losing state on a restart. Also built the quality layer: a promptfoo harness with a multimodal judge that watches the rendered .mp4 and scores it.

PythonInngestFastAPILLM gatewayStructured outputpromptfooLLM-as-judgeEvals

Video Generation — Provider Fallbacks

Video providers go down or reject jobs constantly, so I turned a single-provider call into a multi-provider fallback chain across video-generation vendors that treats a terminal poll response as failure, reorders providers on failure, and skips flaky ones inside a bad time window. Adding a fallback provider dropped the failure rate by more than half and cut average generation time. Added an upscaling pass and an automated suite so the logic can't silently rot.

PythonMulti-providerResilienceFallbacksUpscalingTesting

Deploy — Zero-Downtime Blue-Green

Owned the deploy end to end. Moved the service off managed functions onto self-hosted infrastructure with a blue-green flip — new code comes up on the idle instance, passes a health check, then a reverse proxy swaps to it and the old one drains before retiring, so no in-flight request dies. When CI runners couldn't reach the box, I traced the network path and moved deploys onto a private mesh VPN, which fixed it and closed SSH to the public internet. Took deploys from intermittently failing to consistently reliable.

GitHub ActionsnginxsystemdTailscaleWireGuardBashSSHBlue-Green

Creator Platform — API & Credits

Primary backend engineer on the creator app. Shipped thirty-plus REST endpoints on Lambda for agents, presets, characters, generations, and a filterable central library of everything a user has made. Also built the money side: a credit/billing system that charges atomically as it writes the row and refunds back into the exact source it drew from, wired to card and crypto checkout. Delivered reconciled, dispute-free billing with no accounting drift.

PythonAWS LambdaPydanticPostgreSQLStripeCryptoSolanan8n

Voice Platform — Cloning & Multilingual

Migrated the whole voice catalog to a new TTS/cloning provider and built the cloning flow behind it — hiding internal provider IDs from the frontend and handling duplicate-name collisions cleanly instead of 500-ing. Scaled the catalog from a handful of English voices to 600 across 12 languages with a scrape → clean → seed pipeline, and added social-source voice-clone ingestion with lipsync.

PythonTTSVoice cloningMultilingualPostgreSQL

GPU Workers — Audio & Video Post-Processing

Built two serverless GPU workers on my own. The first strips vocals from a track by running an ensemble of separation models and averaging their stems. The second does automated video inpainting / object removal — detecting the target region (crisp static edges versus moving content), masking it, and inpainting the hole with LaMa before muxing the original audio back. Baked the model weights into the CUDA image so there's no cold-start download.

PythonRunPodCUDADockerOpenCVLaMaaudio-separatorNumPyS3

Autonomous C# Video Engine

Sole developer of the original node-based engine in C# that takes a topic and returns a finished long-form video with no human in the loop — LLM story generation, text-to-image and image-to-video models, LoRA fine-tunes and previous-frame references for character consistency, per-scene voiceover and subtitles, and cloud assembly. The generation pipeline was later rebuilt as a durable Inngest service.

C#.NETText-to-imageImage-to-videoLoRALLMsFFmpeg

Discord Bot — Delivery & On-Chain Pay

Built a Discord bot that lets a user pick which agent to run from a slash command and drops the finished video straight into the channel as a presigned link, backed by a pooled Postgres client and a short-lived cache on the hot lookups. Wired up Solana payment verification through Helius next to Stripe so the agency order flow could take either.

TypeScriptdiscord.jsTelegrafPostgreSQLAWS S3SolanaHeliusStripe

Infrastructure — Terraform & AWS

Own the Terraform that stands the platform up across separate prod and dev environments — serverless functions, storage, event scheduling, private networking, IAM, and the service URLs the webhooks call back on. Provisioned service modules end to end and pushed every provider secret through the pipeline instead of the console, so the whole thing rebuilds from code in either environment.

TerraformAWSLambdaVPCIAMEventBridgeCI/CD

Content Ops — No-Code Showcase Sync

Built a small Python tool so the non-technical team could update the landing-page video showcase themselves. They drop a folder of clips and run one command; it uploads the avatars to Cloudflare Images, writes the metadata to Cloudflare KV, and the live site picks it up with no deploy and no engineer in the loop. Deduplicates avatars by content hash so re-runs don't re-upload anything.

PythonCloudflare ImagesCloudflare KVCLIContent ops

Earlier — Game & XR Development

Before moving to backend/AI, I shipped games and XR as a Unity/C# developer. Built a mobile/WebGL endless runner for a consumer-brand campaign (sole dev — object pooling, difficulty ramp, haptics, REST-backed giveaway) and an interactive Unity video-station kiosk app with an animated tile grid and custom player. Earlier: immersive VR/AR on Quest 2/3 and Jio MR glasses (XR Toolkit, MRTK3, JioMRTK, Oculus SDK), multiplayer WebGL on Photon, and six hypercasual kids' games for BimiBoo.

UnityC#URPWebGLXR ToolkitMRTK3Oculus SDKPhotonDOTween