Hello, I'm
Lekkala Ganesh
AI/ML Engineer |
AI/ML engineer building LLM applications and data-science models in Python — RAG pipelines over private data (chunking, embeddings, vector search), conversational AI with chatbots and voicebots that speak and listen, LangGraph agent orchestration with guardrails, and the FastAPI REST services behind them. Shipped with statistical evaluation, structured logging and CI.
B.Tech CS&E · 500+ automated tests across public repos · Bangalore, India
About Me
I'm an AI/ML engineer who builds the layer where machine learning actually meets users. That means RAG pipelines end to end — chunking, embeddings, and vector search over LanceDB and pgvector — conversational AI in the form of chatbots and voicebots that listen and speak back, and LangGraph agents that retry, self-correct and fail loudly instead of hallucinating quietly.
The other half of my work is data science and evaluation. I build classification and regression models in Python with scikit-learn and pandas, and I measure whether a change actually helped — Cohen's kappa for inter-rater agreement, Wilcoxon signed-rank for paired comparisons, and hand-labelled ground truth scored before any automated judge runs. A 99%-accurate model on an imbalanced set is still a useless model, and knowing why is the job.
All of it ships behind real REST APIs — FastAPI and Express, OpenAPI-documented, JSON payloads validated with Pydantic and Zod, with structured logging, correlation IDs and token-cost tracking so a failure at 2am is diagnosable. I care about evidence over adjectives: my public repos carry 500+ automated tests and CI because "it works on my machine" isn't a claim — it's a hypothesis. B.Tech in Computer Science & Engineering, Parul University. Open to AI/ML engineering and data science roles.
Technical Core
Machine Learning & Statistics
Deep Learning & Model Serving
Generative AI, RAG & Agents
Conversational AI & Voice
Backend & APIs
Data & Storage
Frontend, Testing & DevOps
Experience
Freelance AI/ML & Full-Stack Engineer
Self-Employed · Remote
Conversational AI & LLM Integration
- Chatbots & Voicebots: Built a browser voicebot — speech-to-text → LLM → streaming text-to-speech with barge-in, VAD and a 10-turn context window (VoxAI) — and an ElevenLabs conversational agent that places outbound calls over Twilio.
- Multi-Provider LLM Fallback: Engineered a provider chain (GPT-4o → Gemini → OpenRouter) with graceful degradation, powering real-time SSE streaming and 3D sunlight simulation (SunScope Pro).
AI Training Data & Evaluation
- Rubric Design & Human Annotation: Author scoring rubrics and hand-label model outputs, blind-scored against anchored criteria before any automated judge runs; design adversarial probes for sycophancy, verbosity and position bias.
Backend & Delivery
- PostgreSQL Schema Design: Designed relational schemas with foreign keys, CHECK constraints, composite indexes, and triggers; secured per-user data with Row-Level Security policies and JWT auth on Supabase.
- REST APIs & End-to-End Delivery: Shipped client web apps end-to-end — REST API design with OpenAPI/Swagger, HMAC signing and request-scoped correlation-ID logging, behind React/Next.js front-ends deployed on Vercel — meeting WCAG 2.1 AA with automated Playwright cross-browser testing.
Full Stack Web Development Intern
Slash Mark IT Solutions · Remote, Part-time
- End-to-End MERN Stack Delivery: Delivered 6 full-stack projects across 3 tiers (basic to advanced) on the MERN stack (MongoDB, Express.js, React, Node.js), including a real-time Weather Forecast app with API integration, a URL Shortener REST API (Node.js, Express.js, MongoDB), and a Food Delivery platform with authentication, geolocation, and payment flow.
Machine Learning Intern
Bharat Intern
- Model Development: Built regression and classification models (House Price Prediction, Wine Quality Prediction, Iris Classification) in Python using Scikit-learn and Pandas in Jupyter notebooks — covering data cleaning, exploratory data analysis (EDA), and feature engineering.
- Model Evaluation: Compared estimators using train/test splits, cross-validation, and hyperparameter tuning, selecting final models on held-out evaluation metrics rather than training accuracy.
Projects
Cost-Efficient RAG Service
Chunking → Embeddings → Vector Search → Generation
A full retrieval-augmented generation pipeline: document ingestion, recursive chunking, sentence-transformer embeddings (all-MiniLM-L6-v2, 384-dim, L2-normalised) into a LanceDB vector store, then retrieval and generation — with a second pgvector backend behind the same interface and a written rationale for rejecting FAISS, Qdrant and Chroma.
Agentic Hallucination Guardrail
LangGraph Agent State Machine with Full Tracing
A six-node LangGraph state machine that detects, validates and corrects LLM hallucinations inline — temperature-decay retries (0.7 → 0.25), a three-retry circuit breaker, an 8192-token budget, and Pydantic schema validation on every model output. Each run is traced end to end with latency, retry count and confidence scores, streamed live over SSE.
VoxAI — Real-Time Voicebot
Conversational AI: Speech-to-Text → LLM → Text-to-Speech
A full conversational voice loop — Web Speech API capture streams into Gemini over SSE, and sentence-boundary chunks feed a non-overlapping TTS queue so the reply starts speaking before it finishes generating. Barge-in interrupts mid-sentence, VAD gates the mic, and a 10-turn context window resolves references like "tell me more about the first one". 25+ language translation.
LLM-as-Judge Evaluation Pipeline
Measuring Whether the Judge Can Be Trusted
Quantifies five distinct judge biases — position, verbosity, self-enhancement, sycophancy and score clustering — against 105 human judgments I hand-labelled first, blind, using a self-authored rubric with 1/3/5 anchors. Agreement is measured with Cohen's kappa and paired comparisons with the Wilcoxon signed-rank test, so "the judge got better" is a statistical claim rather than a vibe.
Other Projects
AuraGen
Local-first image generation platform for low-VRAM machines. Async job queue with WebSocket progress streaming and OOM-aware handling; hybrid cloud-burst routing offloads jobs over the local VRAM budget to HuggingFace, falling back to local. SAM2 point-to-edit segmentation.
View Project →ATS Resume Optimizer
Resume-to-job matching engine using scikit-learn TF-IDF vectorization and cosine similarity, with keyword-gap extraction and a typed Pydantic state architecture across ~20K lines of Python. 276 tests under GitHub Actions CI.
View Project →PeanutGuard AI
Offline-first PWA classifying 19 crop-disease classes entirely in-browser through a three-stage ONNX Runtime Web pipeline, with rule-based intent classification driving the advisory chat and IndexedDB → Supabase sync. Telugu and Hindi localisation.
View Project →Urja Meter Ops API
REST API and zero-build console reverse-engineered over a legacy smart-meter portal. Hexagonal architecture, OpenAPI/Swagger documentation, HMAC request signing, and request-scoped correlation IDs threaded through structured logs.
View Project →Lite-Vision
Serves four pretrained ONNX models (SCRFD detection, InsightFace gender-age, FairFace, FER+) through an OpenCV DNN pipeline with multi-model fusion, CLAHE preprocessing and MD5 frame caching. Rate-limited FastAPI, 73 tests.
View Project →AI Interview Trainer
Local-first interview coach on Ollama. Embeddings persisted as SQLite BLOBs via NumPy with hand-implemented cosine-similarity RAG retrieval, an EMA weakness map, and SM-2 spaced-repetition scheduling. Tiered LLM routing with tier-down fallback over SSE.
View Project →2A Construction LLC
End-to-end contractor platform built and shipped for a client — secure financial data handling on RLS-protected backends, React/Next.js front-end, live in production on Vercel.
View Project →AI Stock Portfolio Analyzer
Portfolio concentration analytics — Herfindahl-Hirschman Index, weight distribution, and max-drawdown tracking, with an equal-weight rebalancing backtest against buy-and-hold on simulated price paths. LLM explains pre-computed metrics.
View Project →LinkedIn Post Generator
AI content generation tool for crafting professional LinkedIn posts with tone control.
View Project →VoxAI
Real-time SSE voice streaming assistant with multi-provider AI fallback system.
View Project →SunScope Pro
3D shadow simulation tool for sunlight and comfort analysis with real-time rendering.
View Project →Finance Tracker
End-to-end personal finance app — React client on a JWT-authenticated, RLS-protected PostgreSQL backend with a constraint-backed schema.
View Project →Website Semantic Search
Scrapes a URL and chunks the page along its HTML structure rather than by character count, embeds each chunk and the query with sentence-transformers, then ranks passages by cosine similarity.
View Project →Store Ratings Platform
End-to-end role-based ratings app — React front-end on a Node.js backend, with multi-table JOIN and GROUP BY reporting queries, CHECK constraints, and parameterized SQL across three user roles.
View Project →GoVault
Offline-first travel organizer with GPT-4o powered booking import, itinerary builder, multi-currency expense tracking, and collaborative trip sharing.
View Project →Education
Bachelor of Technology in Computer Science & Engineering
Parul University, Vadodara, Gujarat
Let's Build Something Together
I'm always open to discussing new projects, creative ideas, or opportunities to be part of your vision.