
Tejas Lamba
Backend & AI Engineer
FDE Intern @ CallKaro · open to backend & AI roles
I'm a backend and AI engineer. Right now I'm a Forward Deployed Engineer intern at CallKaro AI, where I build and debug multilingual voice agents handling live customer calls in 20+ languages. Before that, I built embedded AI firmware for STM32 microcontrollers and real-time NPU computer vision pipelines at STMicroelectronics, and shipped fintech backend APIs at Rudescore.
I work across Python, TypeScript, Rust, and C/C++, from native Windows tools (Clementine, a 7.5 MB Rust and Win32 file converter) and reverse-engineering AI agent runtimes (cc-context-diet, cc-subagent-models) and clinical ML systems with SHAP explainability to upstream fixes in microsoft/TokenEconomizer and deepset-ai/haystack, plus npm packages with a few hundred combined stars.
I won the WeMakeDevs x Wispr Flow Challenge with Clementine, and I've taken 1st place at two national hackathons: CodeSlayer 2k25 (DevSphere India, NIT Delhi) with a Pathway AI due diligence copilot, and Fintechstico V7.0 at NSUT in October 2026 with Two Strands, an engine that scores scams and loan default risk for the same customer. I'm currently pursuing a B.Tech in CSE with an AI/ML specialization at Maharishi Dayanand University.
work experience.
Joined the deployment team at a Voice AI startup whose agents handle customer calls and WhatsApp in 20+ languages, owning client onboarding from discovery calls to go-live. Write and iterate on LLM prompts for voice-agent personas and call flows, including Hindi and Hinglish calls, debug live deployments from call logs and transcripts, and build Python scripts and API integrations connecting the platform to client CRMs and workflows.
Engineered low-power embedded AI firmware on STM32U5 microcontrollers with X-CUBE-AI, X-CUBE-MEMS1, and NanoEdge AI Studio, deploying on-device data loggers and anomaly detection models. Built a real-time computer vision system for Tic-Tac-Toe on an STM32MP257 microprocessor running OpenSTLinux, using YOLO11 for state classification, and programmed the game engine and serial communication drivers to coordinate moves with a physical robotic arm.
Built and optimized backend APIs in TypeScript for a high-performance fintech application. Redesigned the PostgreSQL schema for a core feature, cutting query times enough that the team noticed, and used Prisma ORM for type-safe queries with zero-downtime migrations while coordinating with PMs and frontend engineers on feature planning.
Provided IT support across 50+ corporate employees in the General & Administration department, resolving daily tickets covering hardware provisioning, system access control, and network connectivity, while maintaining Active Directory accounts and documenting recurring issues to reduce repeat tickets.
education.
B.Tech in Computer Science Engineering (AI/ML)
Aug 2023 - PresentScore: In Progress
featured projects.








Claude Code skill and zero-dependency Shadow DOM engine that makes 24 LLM cursors rebuild any website from a blank screen.
An installable Claude Code skill (npx skills add TejasLamba2006/swarm-page-builder) and standalone 3,800-line browser engine that snapshots a live page into a Shadow DOM blueprint, wipes the screen, and deploys a swarm of multiplayer LLM cursors to construct every box, text node, border, image, and ruler guide live in front of the visitor.
Zero-click offline file converter and 29-tool AI/media suite in pure Rust and Win32 (~7.5 MB RAM). Winner of the WeMakeDevs x Wispr Flow Challenge.
A native Windows 10 & 11 counterpart to Tangerine for Mac built from scratch in Rust and Win32 GDI/SDF with zero Electron or WebView2 bloat. Sits in the System Tray using ~7.5 MB of RAM, hooks File Explorer and VS Code drags on a dedicated thread, and runs 275 format conversions, 29 studio tools (including pure-Rust U2Netp ONNX background removal, offline Windows.Media.Ocr, and WinRT FaceDetector auto-redaction), and 6 file actions 100% locally.
Binary patcher that cuts Claude Code Turn-1 context bloat by 89%.
An npm CLI and Bun PE binary patcher for @anthropic-ai/claude-code that drops Turn-1 prompt overhead from 211k tokens to 22k tokens via universal gateway ToolSearch, wire-safe MCP schema hydration, and in-place SkillSearch.
Dynamic multi-model routing and per-call context windows for Claude Code subagents.
An npm CLI and runtime patcher for @anthropic-ai/claude-code that unlocks arbitrary gateway model IDs (OpenCode, DeepSeek, Qwen, GLM, Gemini) and per-call context window tags (model[1m]) in the Agent tool.
Real-time YOLOv8n board perception on an STM32MP257 1.35 TOPS NPU with an 11-stage MLOps dashboard.
An end-to-end embedded computer vision system and automated 11-stage MLOps pipeline for the STM32MP257F-DK discovery kit. Reaches 0.995 mAP@0.5 and 100% cell accuracy across 449 test cells while shrinking the model from 6.22 MB (FP32) to 2.68 MB (INT8 NPU binary).
Clinical decision support system for diabetes risk assessment.
A clinical ML system with explainable risk assessment, built for the GDG on Campus Praxis 2.0 Hackathon. Combines an XGBoost classifier with SHAP explainability and a RAG pipeline over DeepSeek V3 for grounded, cited AI responses.
Daily LLM-verified accountability docket for all 543 Indian MPs.
A statically generated public record covering all 543 Lok Sabha Members of Parliament, tracking attendance, MPLADS fund utilization, criminal cases, and campaign promises via an automated daily LLM verification pipeline.
AI photo analysis that matches your face to a Pokemon.
A full-stack web app that analyzes a user's photo and matches their expression/personality traits against 930+ Pokemon characters, using Google Cloud Vision and Gemini AI.
Transform plain text into fancy fonts and styles, instantly.
A React app that transforms text into a variety of Unicode styles and fonts with multi-language support and an instant live preview.
Type-safe npm package connecting Discord.js bots to Lavalink.
A lightweight, type-safe library that connects Discord.js bots to Lavalink audio servers with minimal setup and maximum flexibility. 187 stars on GitHub.
Minecraft Bedwars companion overlay with live stats.
An in-game overlay for Minecraft Bedwars with multi-client compatibility, player stats, and game notifications, used by 200+ players.
Minecraft server admin action monitor with Discord alerts.
An advanced Paper plugin for monitoring administrator actions on a Minecraft server, with Discord webhook integration and pattern-based command matching.
Modern dialog-based login UI for Minecraft's AuthMe plugin.
A modern dialog-based authentication interface for AuthMe on Paper servers, using Minecraft's native dialog system instead of chat prompts.
CLI that keeps your JS dependencies healthy and up to date.
A production-ready CLI for automated JavaScript dependency management, with intelligent update recommendations and semantic versioning safety checks across npm, yarn, pnpm, and bun.
npm wrapper for the Musixmatch lyrics/track API.
An npm package that provides easy access to Musixmatch API endpoints for searching tracks, fetching lyrics, artists, and albums, with automatic authentication.
Real-time ML fraud detection on 284k+ transactions.
An end-to-end ML system for detecting fraudulent credit card transactions in real time, processing 284,000+ transactions at 99.9% accuracy with an interactive Streamlit dashboard.
open source.
TokenEconomizer
6 merged PRs: loop guard, micro-skill learning, Codex tool pruning, Windows hook launcher
hackathons.
No Hands: The Wispr Challenge
@ WeMakeDevs x Wispr Flow · Global (Online)
Clementine: 7.5 MB Native Rust + Win32 File Converter & Local AI Suite
Voice-to-Code & Developer Productivity
- •Won the $100 daily prize at the WeMakeDevs x Wispr Flow Challenge with Clementine, a native Windows counterpart to Tangerine for Mac.
- •Clementine is a zero-click radial file converter written in pure Rust and Win32 with no Electron or WebView2, idling at about 7.5 MB of RAM, roughly 16x lighter than an Electron app.
- •It runs 275 format conversions and 29 local tools, including tract-onnx background removal, offline Windows OCR, and WinRT face and text auto-redaction, all without a single cloud call.
- •Open sourced with a Clementine-Setup.exe installer and a GitHub Releases auto-updater.
Fintechstico V7.0
@ Consilium'26, NSUT · NSUT Delhi
Two Strands: scam detection and loan risk in one engine
PS3: AI-Powered Financial Scam Detection & Loan Risk Management
- •Won 1st place with Team Syntax Terrors at an offline fintech hackathon at NSUT, out of 3,093 registrations.
- •Banks run scam detection and loan risk as separate departments, even though a scam often ends in a missed EMI. Two Strands scores both for the same customer and wires them together: a scam hit raises repayment risk, and debt stress puts the account on mule watch.
- •Live scam scoring with LightGBM and coercion rules, a mule-ring graph with GraphSAGE, and days-to-default prediction with a survival model. Every alert comes with SHAP reasons and a plain case note in English or Hindi.
- •Built on FastAPI, PostgreSQL and Redis Streams, with a React analyst console and a Kotlin Android payment app that reads call state during a payment.
CodeSlayer 2k25
@ DevSphere India · NIT Delhi
Due Diligence Copilot: AI Assistant
Track 2: Due Diligence Copilot (Pathway AI)
- •Won 1st place with Team Deadbits out of 10,562 teams. Our track: build an AI assistant for investors doing M&A due diligence.
- •Built a RAG pipeline using Pathway AI and FAISS that answers questions from live data rooms (financials, contracts, cap tables) and updates in real-time as new documents arrive.
- •FastAPI backend handled document uploads and vector similarity search; the system cited sources and exported summaries for investor review.
technical skills.
certifications.
Career Essentials in Cybersecurity
Microsoft · Comprehensive cybersecurity foundations and best practices
Microsoft Security Copilot (2024)
LinkedIn Learning · Microsoft Security Copilot implementation and usage
Cybersecurity Foundations
LinkedIn Learning · Advanced cybersecurity terminology and threat landscape analysis
JavaScript (Basic & Intermediate)
HackerRank · JavaScript fundamentals and advanced concepts certification
Problem Solving (Basic)
HackerRank · Algorithmic problem-solving and data structures
Professional Skills
IBM · Professional development and collaborative skills
Project-Based Learning and Design Thinking
IBM · Design thinking methodology and project-based learning approaches
github.
blog.
view allTen Merged PRs in Two Weeks: 12,279 False Errors and a Hindi Full Stop
Ten pull requests merged across five open-source repositories in early November: six in microsoft/TokenEconomizer, one in deepset-ai/haystack, and three small cleanups. What broke in each one and how I verified the fix.
Inside Switch: Building 5 Core PRs and Two Multi-Room Agent Systems
What I learned opening 5 core PRs (#1214, #1215, #1216, #1592, #1593, +10,474 LOC and 326 tests across Python/PostgreSQL and the Electron Session Host) on sandbox-quantum/switch for the Flip the Switch Hackathon, and building two multi-room agent workflows (Switch DealDesk and Switch Studio) on top of them.
Building Clementine in Pure Rust and Win32: A 7.5 MB Windows Counterpart to Tangerine for Mac
How I built a zero-click radial file converter and 29-tool local AI suite in pure Rust and Win32 GDI using 7.5 MB of RAM (16x lighter than Electron), solved Windows OLE DoDragDrop hit-testing, and won the WeMakeDevs x Wispr Flow Challenge.
Patching the 249 MB Claude Code Binary: Custom Subagent Models and an 89% Turn-1 Context Diet
How I unpacked the 2,394 embedded Bun modules inside claude.exe to unlock arbitrary gateway models and 1M context tags on subagents (cc-subagent-models), then cut Turn-1 prompt bloat from 211,672 tokens down to 22,433 tokens (cc-context-diet).
Fingerprinting Stealth LLMs With Differential BPE Token Math
How I used 20x amplified BPE token-count deltas, HuggingFace tokenizer.json files, and raw SSE trace IDs to unmask OpenCode's stealth endpoints (beam-free, exo-free, space-bunny-free, cat-jam-free, and big-pickle).