00: Intro
Lakshya Pachisia
AI/ML + Systems
Building systems for efficient, reliable AI.
About
I'm a computer science student at PES University interested in AI/ML systems and the engineering behind them.
I like building things that sit close to the system: efficient LLM inference, reasoning systems, databases, distributed pipelines, and tools that turn research ideas into working software.
Most of my work has involved understanding how something works under the hood, then building or modifying the system myself. Recently, I've been working on LLM inference and multi-agent systems, while also exploring lower-level systems through projects like my own relational database.
I'm particularly interested in problems where machine learning meets systems engineering, especially when performance, reliability, and real-world constraints matter.
Bengaluru, India
01 / Selected Work
A collection of things I've built
Projects across LLM systems, AI/ML, and systems engineering.
- 01
LAKV
Adaptive KV Cache Compression for Multi-Agent LLM Systems
A training-free, layer-aware KV-cache compression framework for Qwen2.5-7B-Instruct. It scores transformer layers using attention importance, offset variance, and effective rank, then applies tiered INT8, INT4, or drop-based compression.
- 2.81x KV-cache compression
- Diagnosed an aggressive INT4 failure mode caused by per-tensor outlier clipping
LLM SystemsPythonPyTorchHuggingFacebitsandbytesSource - 02
AlpacaDB
Relational Database Management System
A relational database engine built from scratch in Python, with its own SQL parsing pipeline, persistent page-based storage, Volcano-style query execution, B-Tree indexing, and cost-based query optimization.
- SQL lexer, parser, AST, executor, optimizer, and storage engine implemented from scratch
- Cost-based selection between sequential and indexed access paths
SystemsPythonSQLB-TreeQuery OptimizationVolcano ModelSource - 03
Logickos
Deterministic LLM Reasoning Verifier
A deterministic alternative to LLM-as-a-Judge that decomposes reasoning traces into a directed acyclic graph and audits their logical structure using graph algorithms: cycle detection, orphan-node identification, and premise-reachability analysis.
- Produces a Reasoning Consistency Score (RCS ∈ [0, 1])
- Evaluation pipeline and dataset loaders for multi-hop and reasoning benchmarks
LLM SystemsPythonNetworkXFlaskSource - 04
Mini-UnionFS
User-Space Union File System
A user-space UnionFS implementation in C using FUSE 3, combining multiple directory layers into a unified filesystem while implementing copy-on-write behavior and whiteout semantics for deletions.
- Implemented in C using FUSE 3
- Copy-on-write semantics with whiteout-based deletion handling
SystemsCFUSE 3Operating SystemsFile SystemsSource - 05
Distributed Image Processing Pipeline
Distributed Image Processing with Apache Kafka
A distributed image-processing pipeline built around Apache Kafka, with a FastAPI master node coordinating worker processes through task, result, and heartbeat topics.
- Kafka-based task distribution between master and workers
- Worker health monitoring through heartbeat messages
Distributed SystemsPythonApache KafkaFastAPIDistributed SystemsSource - 06
Zensense
AI Navigation Aid for the Visually Impaired
A real-time computer vision assistant that combines YOLOv8 object detection with MiDaS monocular depth estimation to infer object proximity and lateral position from a single camera stream, then converts scene understanding into spoken guidance.
- Objects classified as left, center, or right from bounding-box position
- Phi-2 generates concise spoken descriptions of detected objects, with voice commands to activate/deactivate
AI / MLPythonYOLOv8MiDaSOpenCVPyTorchStreamlitWebRTCPhi-2Source - 07
Raksha
AI Conversation Guardian Against Phone Scams
A user-activated Flutter app that listens to speakerphone conversations in real time and flags common social-engineering scam patterns as they happen, speaking a warning aloud when it hears one.
- Live speech-to-text matched against a keyword list for scam language (OTP, urgency, impersonating police or a bank)
- Spoken alerts via text-to-speech when a conversation turns suspicious
AI / MLFlutterDartSpeech-to-TextText-to-SpeechSource - 08
Fed-J
Federated Learning Simulation with Byzantine-Robust Aggregation
A Java simulation of federated learning across honest, noisy, and Byzantine clients, comparing FedAvg and FedProx aggregation strategies with differential-privacy noise injection. Built for an Object-Oriented Analysis and Design course.
- Simulates 5 clients (3 honest, 1 noisy, 1 Byzantine) across 20 training rounds
- A Gaussian-noise decorator adds differential privacy to honest clients, with a live loss-curve visualization of convergence
AI / MLJavaFederated LearningDesign PatternsSource - 09
NFC Payment System
Face Authentication + NFC-Verified Payments
An integrated kiosk payment system pairing face authentication with NFC card verification: a Raspberry Pi controller handles motion detection and NFC scanning, a Python service handles face recognition, and a Next.js frontend ties the ordering flow together.
- A PIR motion sensor triggers face authentication before an NFC tap completes payment
- Raspberry Pi controller, Python face-recognition service, and Next.js frontend operate as one system
SystemsNext.jsPythonRaspberry PiNFCMySQLSource
02: Experience
Where I've worked
Software Development Intern · Visa Inc.
June 2026 – August 2026
- Built SENTINEL, a six-agent GenAI system that converts Jira stories into PySpark validation scripts, with schema-grounded routing across enterprise data domains and deterministic business-rule extraction
- Automated a workflow that previously required roughly a day of manual development
- Schema-grounded routing across 7 enterprise data domains and 135+ columns per domain
- Expanded validation coverage to 18 modular patterns
- Added static analysis and AST-based code repair
- Built a React/FastAPI review interface
PythonPySparkLangGraphLLM APIsTrinosqlglotBackend Intern · Mstack Chemicals
May 2025 – August 2025
- Designed and owned a FastAPI microservice that replaced a Supabase-based backend, delivering 156 REST endpoints across 25+ PostgreSQL models
- Introduced state-machine workflows for enquiries, quotations, and sample requests
- Implemented RBAC at state transitions
- Used feature flags for controlled migration from the legacy backend
FastAPIPostgreSQLSQLAlchemyPydanticSupabaseApp Developer Intern · NDGCS
June 2024 – October 2024
- Built a role-based HR management application with five dashboards, Firebase authentication, and Google Calendar integration for interview scheduling
FlutterFirebaseFigmaGoogle Calendar API
Education
B.Tech in Computer Science Engineering · PES University
2023 – 2027
- CGPA: 9.09 / 10
- CNR Rao Scholarship
- Top 20% of batch across 5 semesters
03: About
A bit about me
Skills
Recognition
- 2nd Place, Inferentia
- 2nd Place, AppGenesis
- Participant, Build India Hackathon 2026
- Deputy Head, Shunya, PES Math Club
- Operations Head, TAMS'24
04: Contact
Let's talk
Reach out about opportunities or collaboration.