pytest-tenantguard
BUILDINGA pytest plugin and AST linter that catches cross-tenant data leaks in multi-tenant SaaS codebases before they ship. MIT-licensed core, with a paid compliance dashboard for teams that need audit trails.
Code that 200,000 network devices depend on. AI agents that handle a million requests a day.
Five years building backend systems, from kernel-level networking at VMware to AI-agentic infrastructure at a fast-moving startup. Currently relocating from Bengaluru to Barcelona for senior and staff backend roles.
0M+
API requests / day
Knowl · hybrid AWS + GKE
0K+
Edge devices served
VMware · VeloCloud SD-WAN
+0%
AI accuracy lift
Language detection agent
0%
Test coverage
Up from 5% · legacy C codebase
01 · ABOUT
about.md
About Abhishek
I'm a backend engineer who likes problems where the constraints are real: limited memory, strict latency budgets, traffic that won't wait for you to scale up. Day to day, that means owning technical decisions, defining API contracts, and mentoring the engineers around me, as much as it means writing code.
Most recently, I was a founding engineer at Knowl, where I built the Python and FastAPI backbone for an AI-native product handling over a million requests a day, including an autonomous agentic system that improved language detection accuracy by 45%.
Before that, I spent four years at VMware on the VeloCloud SD-WAN platform, writing production C and C++ for systems running on 200,000+ edge devices: threat detection, IPv6 prefix delegation, and a testing framework that took a legacy C codebase from 5% to 90% coverage.
I studied Computer Science at IIT Guwahati, and still keep my competitive programming chops sharp. Outside of work, I run local LLMs on a Mac Mini and write about what I find.
02 · EXPERIENCE
Where I've shipped
Founding Engineer
Knowl · Bengaluru, India
Nov 2025 – Jan 2026
- Designed Python/FastAPI backend services handling 1M+ API requests/day on a hybrid AWS EC2 and GKE architecture.
- Built an autonomous AI agentic system for language detection using LLM reasoning and multi-strategy fallback, improving accuracy by 45% through prompt engineering.
- Implemented Redis-based rate limiting and throttling, keeping high-volume APIs stable under P95 load.
- Built scalable REST APIs with Pydantic validation across a reactive microservice architecture.
Member of Technical Staff 2
VMware · Arista (VeloCloud SD-WAN), Bengaluru, India
Jul 2021 – Oct 2025
- Built a network interface tracking system giving real-time visibility into 200,000+ edge devices, from kernel-level data extraction to the user-facing dashboard.
- Delivered Advanced Threat Protection (IDS/IPS, threat intelligence, URL filtering) that mitigated 10,000+ malicious connection attempts a month.
- Designed and implemented IPv6 DHCP Prefix Delegation for dynamic prefix allocation across large SD-WAN deployments with zero-NAT global addressing.
- Led the introduction of a unit testing framework for legacy C code, scaling coverage from 5% to 90% and unlocking full CI automation.
- Built a packet processing latency visualization platform that sped up performance debugging.
03 · PROJECTS
Things I build outside of work
MigrationGuard
BUILDINGA CLI tool that catches backward-incompatible database migrations before they break a zero-downtime deploy, by tracing migration diffs back to the application code that actually depends on them.
Interview Helper
ACTIVEA local RAG application that stores and queries interview experiences with semantic search and metadata filtering, running entirely on local infrastructure end to end: ingestion, indexing, retrieval, and LLM-powered answers.
Librarian
ACTIVEA multi-book conversational AI that ingests PDFs, builds a searchable knowledge base, and answers questions with citations, using hybrid retrieval across vector search, TF-IDF, and cosine similarity.
More on GitHub.
04 · FIELD NOTES
Notes from building with local AI
LOCAL AI · SYSTEMS
The real bottleneck in local AI isn't model size. It's memory.
Running larger coding models on a 16GB Mac Mini, I noticed generation slowing down and the model losing track of earlier context as a project grew. The cause wasn't the model, it was the KV cache eating into available memory as context filled up. A smaller model with more headroom for context can outperform a bigger one on real projects, and I think the next real unlock for local AI is better memory architecture, not bigger weights.
Read the full post on LinkedInBENCHMARK · APPLE SILICON
Benchmarking Qwen3.5 on a Mac Mini against my Claude subscription
Same coding prompt, run across 4B and 9B Qwen3.5 models with and without Apple's MLX backend. Runtime optimization turned out to matter almost as much as model size.
No MLX
13.13 tok/s
MLX
27.71 tok/s
~2.1x faster generation on the 4B model, with total inference time dropping from 4m39s to 1m43s. 4B models now feel genuinely usable for real engineering work.
Read the full post on LinkedIn05 · STACK
What I build with
Languages
PythonTypeScriptCC++Java
Cloud & Infra
AWS (EC2, S3, SQS)GKETerraformDockerKubernetesKEDA
AI & LLM
AI Agent DevRAGPrompt EngineeringLangChain
Frameworks
FastAPIFlaskNode.jsExpress.jsSpring
Data
PostgreSQLMySQLRedisMongoDBChromaDBElasticsearch
Practices
CI/CDAgileUnit TestingMicroservicesTCP/IPNetwork Security
Retrieval Augmented Generation (RAG)
April 2026 · Present
Claude Code: An Agentic Coding Assistant
May 2026
AWS Storage
June 2025
06 · CONTACT
Open to senior and staff backend roles in Barcelona and across Europe.
I'm relocating from Bengaluru and ready to start the visa sponsorship process. If you're hiring backend or AI-systems engineers, I'd love to talk.