context.md
For recruiters, collaborators, or any AI assistant evaluating fit. Pull the raw file or paste it into any AI assistant.
About
I'm Arnav Thakur. Software engineer. I build backend systems, infrastructure, and AI platforms. I've spent time across production platform engineering at Trellix, published ML research during college, and independently shipped platforms and systems like k8s-agent, Hookd, and RecapGen.
I'm drawn to problems at the intersection of systems, AI, and reliability. I enjoy going deep on how things work, reading source code, and building tools people actually use.
Areas of Interest
Skills
Languages
AI Infrastructure
Backend Engineering
Platform Engineering
Security
Developer Tools
Research
Published during undergraduate study at Amrita School of Engineering, Bangalore. This is where my ML foundation was built.
5 papers · Deep Learning, Computer Vision, Interpretable ML
Professional Experience
Trellix, Platform Engineering (July 2025 - current)
Embedded on the Platform team working alongside Staff Engineers and Senior Staff Engineers on production infrastructure, security systems, Kubernetes, and AI deployment.
Platform Engineering
SupportRelay: PowerShell automation now live in production. Collaborated with support staff to identify requirements and surface field engineer pain points. Jumpstart DNS refactor to cleanly separate IPv4 and IPv6 parameters. Cross-platform build pipeline debugging.
Security Engineering
Hardened DeviceAPI TLS: replaced InsecureSkipVerify with proper X.509 validation, custom CA pools, hostname verification. Mapped FIPS and FedRAMP compliance requirements into code-level decisions.
AI Infrastructure (Project Trella)
Air-gapped RAG chatbot for Trellix security appliances. Six-service Helm umbrella chart on RKE2, zero external network dependency at runtime. FastAPI orchestrator with three-branch decision tree. Intent classifier (TF-IDF + logistic regression, 7 classes) that short-circuits roughly half of user turns, enabling the full stack on a 4-core CPU-only node. Hot-reload mechanism: nightly retrains swap the in-memory classifier with no pod restarts, no dropped requests. Offline Ollama inference, weights baked into container image. Single air-gapped install artifact at about 78% of available compute. Structured audit logging piped to stdout for the customer's log aggregation pipeline. Authored architecture doc and led internal walkthroughs with platform engineering staff and senior staff engineers.
Kubernetes (RKE2)
Initial POC for bringing RKE2 to the base platform, working closely with senior engineers on server-agent configuration and HA design. Resolved silent packet drops via hand-coded iptables rules. Caught a kube-proxy conflict that enabled eBPF in the final delivery. Kubernetes Secrets for OIDC rotation.
Configuration Management
Collaborated with staff engineers to understand the mgmtd framework before writing md_modules for secret updates and configuration persistence. Config nodes, side effects, appliance state across restarts.
Projects
k8s-agent
An autonomous, Kubernetes-native AI SRE and incident response platform. Bridges ReAct and native Function Calling across a strict 12-tool diagnostic registry with domain entity-overlap episodic memory. Proactive warning watchers stream K8s anomalies through cooldown deduplicators, while write mutations are strictly safety-gated behind loop-boundary operator approvals. Python, FastAPI, PromQL/LogQL, Helm. github.com/arnav7thakur/k8s-agent
reusecv
Career-as-Code knowledge base and compiler. Replaces disparate career artifacts with a single YAML/Markdown repository. The Go CLI and REST API compile tailored resumes, PDFs (via tectonic), and portfolios on demand using semantic filtering via Gemini embeddings. Deployed via Docker to AWS EC2 using Terraform, with full CI/CD release automation and structured observability. github.com/arnav7thakur/reusecv
Hookd
Started from a personal frustration: saving content is easy but finding it again is not. Existing tools index by title or keyword, which is not how people actually remember things. Hookd captures a short intent note at the moment of saving and uses a two-tier hybrid search engine (zero-cost regex + asymmetric vector retrieval fallback) with cosine similarity re-ranking and stale-while-revalidate local caching. Flutter PWA and Android client, FastAPI backend, MongoDB Atlas, Vercel and Render. gethookd.vercel.app
RecapGen
Turns a folder of photos and a song into a video where cuts actually feel musical. Most tools cut mechanically on a fixed BPM. RecapGen uses librosa to compute onset strength envelopes and merge beat tracking with a JIT pre-compilation container warmup cutting cold starts by 91%. Heavy DSP and parallel multi-core FFmpeg transcoding run in isolated Modal serverless containers to bypass edge gateway timeout limits. Python backend, Vite frontend, Vercel. recap-gen.vercel.app
This is who I am
I've built across production infrastructure, AI systems, independent products, and published research. Still exploring, still going deeper. The interesting problems tend to sit at the intersection of things.
If something here resonates but there's a gap between what you need and what you see — let's talk anyway. I pick things up fast, work independently, and tend to build solutions rather than wait for them. I'm more interested in the problem than in checking every box on a list.
Paste arnav7t.tech/llms.txt into any AI assistant for a more interactive session.