Sheet 01 — General arrangement

Lovekesh Anand, backend engineer. Research. Design. Build.

I'm Lovekesh Anand, a backend engineer. I research and design the system first, then build it. Lately: AI infrastructure.

View the sheets ↓Email me

Role: Backend engineer
Core: Node · FastAPI · Postgres
AI infra: llama.cpp · CUDA · RAG
Record: 4 podiums · 1st @ Microsoft
Status: Open to hire
Résumé: View ↗

Sheet 02 — Method

I design the system before I write the code.

01
Research
Papers, constraints, prior art.
02
Design
Data flow, failure modes, trade-offs.
03
Build
Boring, tested, measured.
04
Ship
Deploy, monitor, write it up.

Sheet 03 — Systems

Six systems. Drawn first, then built.

03.1Agency operations engineLive ↗Source ↗

Fig. A — ArchitectureSwipe →

Winner · Automate India

Kairos

Your code is the engine. Notion is the interface.

Problem: Agency work is scattered across WhatsApp, Gmail and Meet, and someone has to turn it all into tasks by hand.
Design: An event pipeline structures every message with an LLM router, waits for approval in Notion, then executes it.
Result: 1st place, Notion track, Automate India at Microsoft. Built in 8 hours.

Stack — Node.js / Express / Notion API / OpenRouter / WhatsApp / Gmail OAuth

03.2LLM memory researchSource ↗

Fig. B — ArchitectureSwipe →

Nyxen Memory

My AI doesn't forget anymore.

Problem: LLMs forget once the context window fills, and RAG wastes tokens re-pasting old text.
Design: Conversations become a knowledge graph, and relevant facts are injected straight into the model's attention as KV-cache.
Result: Recalls decisions from hours earlier with zero extra prompt tokens, fully local.

Stack — Python / PyTorch / CUDA / Knowledge graphs / KV-cache

03.3Local AI coding engineSource ↗

Fig. C — ArchitectureSwipe →

Epsilon IDE

Three local models, routed by how hard the task is.

Problem: A real multi-model coding assistant normally needs a server, not a laptop GPU.
Design: Prompts are scored 1–10 and routed to a 1.5B, 7B or 33B model. A six-agent pipeline pulls context with TF-IDF + sqlite-vec search.
Result: 30–80 tok/s fully offline, using flash attention, an INT8 KV cache and sparse top-64.

Stack — Python / CUDA / llama.cpp / FastAPI / sqlite-vec / TF-IDF

03.4Self-learning browser agentSource ↗

Fig. D — ArchitectureSwipe →

2nd · SLAB Hackathon

StockSentinel

Let the agent watch and reason. A human makes the final call.

Problem: Fully autonomous trading bots remove human judgement, which is risky when LLMs hallucinate.
Design: A browser agent learns each site's DOM, a local Qwen 2.5 7B reasons over the signals, and every trade needs Telegram approval.
Result: 2nd place at the SLAB Hackathon (webcmd, MAIT).

Stack — Python / Browser agents / Scrapling / Qwen 2.5 7B / Telegram

03.5On-device voice AILive ↗Source ↗

Fig. E — ArchitectureSwipe →

Part of HOL Guard

Kelados

Speak. It's written.

Problem: Hosted voice AI charges per character, uploads your audio and caps your usage.
Design: Speech-to-text and text-to-speech run on your own device, with a cleanup layer that follows spoken commands.
Result: Unlimited, private, $0 server cost. Now part of the HOL Guard ecosystem.

Stack — TypeScript / WebGPU / Whisper / Moonshine / npm SDK

03.6Offline legal AISource ↗

Fig. F — ArchitectureSwipe →

2nd · Matrix 3.0

NyayaAI

Indian legal drafting that never leaves your machine.

Problem: Law firms can't send client files to cloud AI, and the IPC-to-BNS switch broke older references.
Design: A fully local pipeline: OCR, then a llama.cpp model grounded in an IPC→BNS map and precedent search.
Result: 2nd place at Matrix 3.0 (36 hours). Court-ready drafts on a 4 GB laptop.

Stack — Python / FastAPI / llama.cpp / SQLite / Tesseract / React

Appendix: other drawings

All 40+ repos ↗

Sheet 04 — Field record

7+ hackathons. 4 podiums. Built under a clock.

7+
Hackathons entered: 4
Podium finishes: 1st
Automate India @ Microsoft: 40+
Open-source repos

Lovekesh at Automate India — Fig. 1Automate IndiaHackBriven · Microsoft, NoidaNotion track winner. Built Kairos in 8 hours.

Lovekesh at Ctrl Alt Hack — Fig. 2Ctrl Alt HackHackBriven · Microsoft, NoidaPodium with Dua Saeed.

Lovekesh at SLAB Hackathon — Fig. 3SLAB Hackathonwebcmd · MAIT, DelhiStockSentinel, a human-gated trading agent.

Lovekesh at Matrix 3.0 — Fig. 4Matrix 3.0IITM Janakpuri · 36 hrsNyayaAI, offline legal AI.

Fig. 5 — Open sourceKelados joined the HOL Guard ecosystem.feat: add Kelados TTS · mergedAll posts on LinkedIn ↗

Sheet 05 — Specification

Backend first. AI infra by obsession.

Bill of materials

Item	Component	Parts
01	BackendNode.js, Express, FastAPI, PostgreSQL, MongoDB, Prisma, WebSockets, JWT, RBAC	Node.js, Express, FastAPI, PostgreSQL, MongoDB, Prisma, WebSockets, JWT, RBAC
02	AI infrastructurellama.cpp, CUDA, RAG, KV-cache, agents, Whisper, local inference	llama.cpp, CUDA, RAG, KV-cache, agents, Whisper, local inference
03	Cloud & DevOpsAWS EC2, VPC, Jenkins CI/CD, PM2, Docker, Nginx, OpenVPN, Linux	AWS EC2, VPC, Jenkins CI/CD, PM2, Docker, Nginx, OpenVPN, Linux
04	InterfacesNext.js, React, TypeScript, Tailwind	Next.js, React, TypeScript, Tailwind

General notes

1.Research before code. Know why the obvious answer fails.
2.Draw the system. If it can't be drawn, it can't be built.
3.Measure, then optimise. Write down what broke.

Revision history

Rev	Role	Note
C	IT InternEaseInfra · Dec 2025 – now	AWS EC2/VPC, OpenVPN, PM2, Jenkins CI/CD, and production debugging.
B	Technical HeadNexturn	Ran the tech side of the internship cell.
A	B.Tech, Computer ScienceMDU · Class of 2027	Aug 2023 – May 2027.

Lovekesh Anand, backend engineer. Research. Design. Build.01Research.02Design.03Build.