Hey there!
Sourik Adhikary

Founding Engineer | Backend & ML Systems | GenAI Infra

Shipping sub-50ms backends, multi-GPU pipelines, and local-first AI engines from 0 to 1

Experience across:
Foosh AI Amazon Razorpay Tekion
Scroll Down

About Me

Sourik Adhikary
That's me! (Click to bonk)
BONK!
Hey!
Clicks: 0
⚡
🎯
🚀

Hey there! 👋

I'm a Staff / Senior Backend & ML Systems Engineer with 7+ years of experience designing high-throughput distributed backends and production GenAI infrastructure.

Focus Areas:

  • Multi-GPU Inference & ML Infrastructure
  • Storage Engines, Queues & Local-First AI
  • Multimodal Pipelines (Vision, Video & Audio)
  • Low-Latency Distributed Backends
  • High-Reliability Payment & Ledger Rails
0 Years+ Exp 📅
0 M+ AI Gens 🧬
TEKION
Foosh AI

My Journey

🧠

Independent Consultant (Backend & ML Systems)

Remote (US & Global) Jan 2022 - Present
  • Stealth AI (Founding Architecture): Architected a 6-phase persistent memory pipeline with Gemini structured extraction (95% accuracy) and hybrid Pinecone BM25+dense retrieval with temporal decay, cutting semantic check overhead by 95% and lifting retrieval from 62% to 81%.
  • Creator Fintech & Ledgers: Designed core ACH payout, KYC, and lending rails supporting 400+ creator agencies, plus automated double-entry reconciliation engines syncing with QuickBooks and Xero.
PythonGeminiPineconePostgresAWS
🚀

Founding Engineer (Backend & ML Infra)

Foosh.AI Aug 2024 - Dec 2025
  • First engineer; owned backend architecture, distributed GPU orchestration, and product execution from prototype to production.
  • Architected distributed Flux 1.1 training/inference pipelines across Modal and DataCrunch plus latent interpolation upscaling, powering 2M+ enterprise generations.
  • Designed Foosh v2, an async DAG workflow engine chaining image, video, and lip-sync models with automated fallbacks.
PythonPyTorchCUDAModalFastAPI
🔧

Software Engineer (SDE 2)

Tekion Corp May 2021 - Dec 2021
  • Re-architected core feature definitions and database queries, lifting availability from 99.0% to 99.99% and cutting p95 latency by 90% (600ms → 50ms).
  • Designed cross-cutting context-preserving Notes infrastructure across microservices, cutting cross-page context switches by 70%.
JavaSpring BootMongoDBGCP
💳

Software Development Engineer

Razorpay Nov 2020 - Mar 2021
  • Decomposed Tier-1 payment routing engines from a legacy PHP monolith into high-throughput Go microservices with zero downtime.
  • Integrated direct bank gateway adaptors into smart UPI routing, improving transaction success rates under peak loads.
GoKafkaRedisPayments
📦

Software Development Engineer (Full-Time & Intern)

Amazon Feb 2019 - Oct 2020
  • Engineered a high-availability retail merchant ingress microservice on AWS maintaining sub-50ms p99 latency at scale.
  • Migrated a mission-critical cryptographic key service and deployed anomaly alerts preventing merchant drop-offs.
  • Built a distributed merchant configuration crawler as an intern, recovering $10k+ monthly TPV and earning a return offer.
JavaAWSDynamoDBDistributed Systems

Featured Work

🎬

Vindex

Local-first video search & AI dataset engine for MCP agents

Hover / tap to flip ↻

Compiles raw video into deterministic bundles via an 8-op content-addressed Merkle DAG over SQLite-WAL, crash-only lease reaper, warm Whisper/CLIP sidecar, and hybrid sqlite-vec + FTS5 RRF search.

10.4x Cache
🧬

Foosh AI GenAI Infra

Multi-GPU Flux orchestration & async multimodal workflow engine

Hover / tap to flip ↻

Founding engineer owning backend & ML infra. Built Flux 1.1 training/inference across Modal and DataCrunch, custom latent interpolation upscaling, bulk moodboard templating, and Foosh v2 DAG workflows.

2M+ gens
🦀

TraceGit

Content-addressed storage engine & deterministic replay for AI agents

Hover / tap to flip ↻

On-disk content-addressed object store in Rust with Blake3 Merkle DAG hashing, zstd-compressed packfiles, AST/event state diffing, and deterministic session replay with mocked I/O for incident post-mortems.

Rust Engine
📦

Shannonic

Near-entropy-optimal lossless compression for ML tensors

Hover / tap to flip ↻

Reimplemented an MLSys 2026 paper as a Rust+Python hybrid using DP range partitioning and streaming rANS coding. Reached within 0.3–2.9% of Shannon's limit and beat zlib on 20/21 ResNet-18 layers.

1.55x Ratio
📺

Tracer

Autonomous sponsorship logo auditor for sports broadcasts

Hover / tap to flip ↻

Zero-training open-vocabulary logo detection running Gemma 4 26B-A4B MoE via MLX on Apple Silicon, computing four-factor QoE scores, cropped proof frames, and interactive HTML heatmap reports.

Open Vocab
📻

Infinite Echo

Autonomous gapless AI radio station on Apple Silicon

Hover / tap to flip ↻

Local AI radio station orchestrating ACE-Step 1.5 music generation via MLX and Kokoro TTS with a look-ahead scheduler that masks generation latency for nonstop playback with zero buffer underruns.

0 underruns

Open Source Arcade

🔥
Loading graph...
🔥
⏳
Fetching repos from GitHub...

Dev Log

📰
Loading posts...

Tech Stack

Languages & Systems

Python
Rust & PyO3
Go
Java

ML Infra & Inference

PyTorch
NVIDIA CUDA
vLLM
Apple MLX & Metal

Distributed Backends

FastAPI
Spring Boot
AWS
Apache Kafka

Production Arsenal & Ecosystem

SGLang
Triton
Modal
Hugging Face
Gemini
Pinecone
FFmpeg
OpenCV
C / C++
SQLite
PostgreSQL
Redis
DynamoDB
MongoDB
RabbitMQ
Google Cloud
Node.js
TypeScript
React
Docker
Kubernetes
Git

Let's Connect!

Building hard ML systems or distributed infrastructure? Need a founding/staff engineer?

Or just want to chat about GPUs, local-first AI & storage engines? 🚀

I'd love to hear from you!

📍 Bangalore / Kolkata / Remote