Skip to content
View 57Ajay's full-sized avatar
🎯
Focusing
🎯
Focusing
8 followers · 18 following

Achievements

Achievement: Pair Extraordinairex2Achievement: QuickdrawAchievement: Pull SharkAchievement: YOLO

Achievements

Achievement: Pair Extraordinairex2Achievement: QuickdrawAchievement: Pull SharkAchievement: YOLO

Highlights

  • Pro

Block or report 57Ajay

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
57Ajay/README.md

Ajay Upadhyay, AI engineer building production AI systems, from model weights to Kubernetes. A trace-style chart of the stack he works across: AI agents in production (10k+ tasks a day), real-time voice, LLM training, inference and serving, and OpenTelemetry contributions.

LinkedIn: upajay X: @_upajay Email: 57ajay.u@gmail.com

I build AI systems that hold up in production, along with the infrastructure that keeps them running. My work spans the whole path from model to user: agents and real-time voice AI, LLM training and inference, Kubernetes platforms, and the observability that ties it all together.

Right now I'm building an LLM stack from scratch, from pretraining in PyTorch to inference in Rust and serving in Go. Those repos are pinned below.

Production highlights

  • Agentic automation at scale. Architected a browser-agent platform (Gemini on Vertex AI, browser-use, Patchright) that runs high-volume operational workflows like traffic-challan settlement and border-tax filing end to end: 10,000+ tasks a day, 500K+ completed.
  • Unit economics. Took per-task handling from ~10 minutes to under 2.5 minutes at ~₹0.05 per task, the daily throughput of ~100 human agents at under 1% of the cost.
  • Support automation. Paired the agent pipeline with fine-tuned reply-drafting models, cutting ticket resolution from 1–2 days to 10–15 minutes.
  • Real-time voice AI. Built a Hindi/Hinglish voice assistant on LiveKit and Gemini that handles live calls, with self-hosted STT/TTS to cut latency and per-call cost versus managed speech APIs.
  • Platform ownership. Built the AI and automation stack from zero for a platform serving 100K+ drivers, and run it as the sole DevOps and MLOps engineer. Migrated production to a custom-autoscaled edge-plus-worker architecture on GCP, now on GKE, and own deployment, autoscaling, reliability, and observability end to end.

Expertise

AI / ML engineering

  • LLM pretraining in PyTorch: DDP, bf16, torch.compile
  • Fine-tuning and alignment: SFT, DPO
  • Tokenizers, data curation, and evals
  • Agentic systems and real-time voice AI

Systems engineering

  • LLM inference internals: GQA, RoPE, KV caching
  • Paged KV cache and continuous batching
  • Q8/Q4 quantization with AVX SIMD kernels in Rust
  • Benchmarking and logit-level verification

MLOps and infrastructure

  • GCP: GKE, Compute Engine, Vertex AI
  • Kubernetes, Docker, Terraform, Linux, CI/CD
  • Autoscaling on queue depth and TTFT SLOs, GPU scale-to-zero on spot
  • Observability: OpenTelemetry, Prometheus, Grafana

Software engineering

  • Rust, Go, Python, TypeScript
  • Concurrent, high-throughput pipelines and worker fleets
  • Servers and APIs: axum, gRPC, OpenAI-compatible endpoints
  • Language internals: OTTL grammar to evaluation

Open source

Merged pull requests to OpenTelemetry (live count)

Contributor to the OpenTelemetry Collector (CNCF), across the core collector and collector-contrib:

  • Built the Coalesce() OTTL converter end to end, from grammar to evaluation
  • Changed the default error mode of the widely used transform processor
  • Added a gRPC user-agent override to the collector's core configuration layer
  • Also merged work across receivers (Kafka, Google Cloud Pub/Sub, host metrics), exporters (AWS S3, debug), and auth/encoding extensions

Tech stack

Rust, Go, Python, TypeScript, PyTorch, Google Cloud, Kubernetes, Docker, Terraform, Linux, GitHub Actions, Prometheus, Grafana, Git

OpenTelemetry Gemini Vertex AI LiveKit

GitHub stats

GitHub stats for 57Ajay

Based in Gurugram, India, and open to remote roles or relocation. Email is the fastest way to reach me: 57ajay.u@gmail.com

Pinned Loading

  1. c_learn c_learn Public

    learning in c

    C 1

  2. distPubSub distPubSub Public

    Simple Redis like DB & distributed pub sub

    Go 1

  3. go_learn go_learn Public

    Go 1

  4. random_art random_art Public

    This generates random images and arts, CURRENTLY IN DEVELOPMENT

    C++ 1

  5. gem gem Public

    A CLI based AI assistant powered by google gemini

    TypeScript

  6. homefindr homefindr Public

    JavaScript 1