Hire Me →
Available 2026 AI/ML Engineer Mumbai, India

I build AI that earns human trust.

I get genuinely excited about AI that closes real gaps — not demos, not dashboards. Real systems. Real humans. Real stakes. Six shipped projects. One thread: building things people actually trust.

See Featured Work Chat with My AI Assess Fit 📄 Download Resume
30-Second Brief
K Sai Sovit
AI / ML Engineer · B.Tech AI
LocationMumbai, India · Remote OK
CurrentDSJ Keep Learning — AI/ML Eng. (May 2025–)
Core StackLangGraph · LangChain · RAG · PyTorch · DenseNet · MTCNN · Transformers · WhisperX · Redis
AvailableOpen — AI/ML, RAG & Agent consulting
Contactksaisovit@gmail.com
// Three things I'm proud of
96.7%Pneumonia detection sensitivity — clinical-grade, trusted by radiologists
3×RAG systems in production — WhatsApp/Email/SMS lead agent, financial analyzer, query bot
+63%Radiologist acceptance increased through Grad-CAM explainability
Say hello → ksaisovit@gmail.com
96.7% Sensitivity 97.1% AUC Score +63% Radiologist Trust LangGraph Multi-Agent in Prod. 3× RAG Deployed 8 Algorithms Benchmarked 92% Attendance Accuracy −18% False Positives 96.7% Sensitivity 97.1% AUC Score +63% Radiologist Trust LangGraph Multi-Agent in Prod. 3× RAG Deployed 8 Algorithms Benchmarked 92% Attendance Accuracy −18% False Positives
Deep Dive

Internship Projects
Architectural Decisions★

Lead Scoring Pipeline
Asynchronous Exponential Decay Redis Queues CRM Integration
Situation
Manual lead scoring was slow, inconsistent, and didn't account for engagement decay over time. Sales teams couldn't prioritize effectively.
Approach
Built an asynchronous lead scoring system using Python and Redis queues. The model incorporates communication patterns, engagement metrics, and behavioral analysis. Exponential decay algorithm weights recent interactions more heavily, ensuring scores reflect current interest. Integrated into the existing CRM for real‑time updates.
Result
Automated lead prioritization, reduced manual effort by 70%, and improved sales focus on high‑intent leads. The exponential decay captured recency better than static scoring.
Key Lesson
Timeliness matters more than total engagement. A lead that interacted yesterday is far more valuable than one from three months ago, even with the same total activity.
Call Scoring Pipeline
WhisperX pyannote.audio Redis Job Queue Rubric Scoring
Situation
Call recordings were reviewed manually for quality assurance — time‑consuming and subjective. No standardized way to measure agent performance or compliance.
Approach
Built a real‑time audio pipeline: WhisperX for high‑accuracy transcription, pyannote.audio for speaker diarization (separating agent from customer). A distributed Redis job queue handles asynchronous processing, scaling across multiple calls. A custom rubric‑based scoring engine evaluates call transcripts against predefined criteria (e.g., greeting, problem understanding, closing).
Result
100% automated call scoring, consistent across all agents. Feedback turnaround dropped from days to minutes. Redis queue allowed horizontal scaling during peak call volumes.
Key Lesson
Speaker diarization is the hardest but most critical part. Without knowing who said what, scoring is meaningless. WhisperX + pyannote together gave the best accuracy vs. latency trade‑off.
Query Chatbot
RAG LangChain OpenAI GPT‑4 PostgreSQL SQL Injection Prevention
Situation
Ed‑Tech CRM had unstructured data across documents, support tickets, and internal knowledge bases. Support teams spent hours answering repetitive queries. Direct SQL access posed security risks.
Approach
Developed a RAG‑powered chatbot using LangChain and OpenAI GPT‑4 API. Retrieval from vectorized internal docs and a PostgreSQL database. Built a secure SQL query validator that parses and sanitizes any generated SQL before execution, preventing injection attacks. The system can answer both document‑based questions and pull structured data from the CRM.
Result
Reduced support load by 40%, response times from hours to seconds. The SQL validator blocked all injection attempts in testing, making the system safe for production.
Key Lesson
Generative SQL is dangerous without a strict validator. We implemented a whitelist‑based query parser that only allows SELECT statements on specific tables — no ALTER, DROP, or JOINs beyond approved ones.
Selected Work

Six Problems
I Got Obsessed With02

01
End-to-EndMost In-Demand '26Multi-AgentRAG
Lead Nurturing Agent
LangGraph · Production · DSJ Keep Learning
3 specialist agents orchestrated via LangGraph state graph. The architecture enables non-linear conversation branching across WhatsApp, Email, and SMS — impossible with a standard chain. RAG via SentenceTransformers + Qdrant for context retrieval.
SituationManual lead qualification across 3 channels — inconsistent, slow, couldn't scale
ApproachLangGraph 3-agent (Qualification / Technical / Pricing) + RAG (SentenceTransformers + Qdrant)
ResultFully automated qualification; real-time personalized responses across all 3 channels
LessonGraph state enables non-linear conversation flows that chains fundamentally cannot handle
LangGraphQdrantSentenceTransformersPython
02
Clinical GradeComputer Vision
Pneumonia Classification
96.7% Sensitivity · 97.1% AUC · +63% Trust
Full case study above ↑ — DenseNet-121 + Bayesian Dropout + Grad-CAM. The explainability layer drove clinical adoption more than raw accuracy.
Key insightClinicians need to explain decisions to patients — a black-box 94% model fails this. Explainability + uncertainty quantification were the real product.
DenseNet-121Grad-CAMBayesian DropoutPyTorch
03
Prod. RAG
LLM Financial Analyzer
Hours → Minutes
RAG on PDF annual reports. Domain-specific chunking strategy for financial structure. LangChain + Pinecone + OpenAI.
SituationAnalysts spending hours per PDF report manually extracting structured insights
LessonRAG + proper chunking beats fine-tuning for doc QA — faster to deploy, easier to update
LangChainPineconeOpenAI
04
Optimization
VRP — 8 Algo Benchmark
Hybrid IP+GA Optimal
GA, SA, ACO, Dijkstra + 4 hybrids. Systematic empirical benchmarking vs theoretical preference.
LessonNo single algorithm dominates all metrics. Empirical benchmarking over theoretical preference — the methodology is the contribution.
PythonDEAPOR-Tools
05
NLP Research
Sanskrit Transformer
Built From Scratch
Transformer from primitives + custom word2vec on scraped corpus. Low-resource NLP is a data strategy problem.
LessonThe challenge wasn't the model — it was corpus strategy. Low-resource NLP is fundamentally a data acquisition problem.
PyTorchWord2VecBeautifulSoup
06
End-to-EndCV · Deployed
Automated Attendance System
92% Accuracy · Live
MTCNN + FaceNet + SVM + geolocation + Streamlit. From camera input to real-time attendance log with fraud prevention. The philosophy: a deployed 92% creates more value than a 99% notebook.
MTCNNFaceNetSVMStreamlitOpenCV
SituationProxy fraud and manual attendance with no geographic verification
ApproachMTCNN face detection + FaceNet embeddings + SVM classifier + geolocation gate + Streamlit UI
Result92% accuracy on real dataset. Full pipeline camera → attendance log, fraud-resistant
LessonShipping is the final feature. A deployed 92% system beats a 99% notebook that never runs in production.
Recent Internship Projects

Shipped at DSJ Keep Learning
Production Systems★

Lead Scoring Pipeline
Asynchronous · Exponential Decay
Developed an asynchronous lead scoring model that incorporates communication patterns, engagement metrics, and behavioral analysis with exponential decay algorithms for sales optimization. Integrated into the existing CRM for real‑time scoring.
Python Redis FastAPI Scikit‑learn
Call Scoring Pipeline
Real‑time Audio · Speaker Diarization
Implemented real‑time audio transcription using WhisperX with pyannote.audio for speaker diarization. Built a distributed Redis job queue for asynchronous processing and integrated a custom rubric‑based scoring system for agent evaluation.
WhisperX pyannote.audio Redis Python
Query Chatbot
RAG · PostgreSQL · SQL Injection Prevention
Developed a RAG‑powered chatbot for an Ed‑Tech CRM using LangChain and OpenAI GPT‑4 API. Designed and implemented a secure PostgreSQL query validator to prevent SQL injection and ensure data integrity.
LangChain OpenAI PostgreSQL RAG
Who I Am

Hi — I'm Sai.
Let me explain.03

Somewhere between a clinical AI project and a room full of sixth-graders learning maths, I figured out what actually drives me: closing the gap between people who have access to good tools and people who don't. That's not a mission statement — it's the thread connecting everything I've built.

"A deployed 92% system creates more value than a 99% notebook. Shipping is the final feature."

My pneumonia project started as a classification task and ended as a lesson in human trust. The +63% radiologist acceptance didn't come from better accuracy — it came from making the model's reasoning visible through Grad-CAM. Clinicians could point to the exact lung region the model flagged. Shared visual language. Suddenly it wasn't "the AI says so" — it was a conversation between doctor and tool.

I volunteer-teach 6th and 7th graders at the We Can We Will Foundation — English, maths, the basics. The access problem I try to solve with code, I also try to solve in person. B.Tech AI at VijayBhoomi University, leading the Alterlights Project, and spending way too much time debating architecture decisions with people who care about them as much as I do.

I'm looking for a team that ships things, debates the hard tradeoffs honestly, and still asks “but does it actually help someone?” after the model hits production.

Sai
K Sai Sovit
AI / ML Engineer · VijayBhoomi University
Emailksaisovit@gmail.com
Phone+91 93371 99404
GitHubksaisovit
LinkedInksaisovit
StatusOpen to hire · AI/ML and Agent roles
// How I think about problems
01
Trust over accuracy — a model people actually use beats a better one they won't.
02
Architecture before tools — LangGraph vs LangChain is a state management decision, not a preference.
03
The question matters as much as the answer — eight VRP benchmarks taught more than one optimal result.
04
Shipping is the final feature — but only if it helps someone when it lands.
Timeline

Where I've
Learned Most04

May 2025 — Present
DSJ Keep Learning
AI / ML Engineer
  • Lead Nurturing Agent — LangGraph 3-agent + RAG. Replaced manual triage across WhatsApp, Email, SMS.
  • Lead Scoring Pipeline — Asynchronous scoring with exponential decay and behavioral analysis, integrated into CRM.
  • Call Scoring Pipeline — WhisperX transcription + pyannote diarization, Redis queue, rubric‑based scoring for agent evaluation.
  • Query Chatbot — RAG + PostgreSQL query validator for Ed‑Tech CRM, preventing SQL injection.
2021 — Present
VijayBhoomi University
B.Tech AI · Research
  • Pneumonia Classification — DenseNet-121 + Bayesian Dropout + Grad-CAM. 96.7% sensitivity, +63% clinical acceptance.
  • LLM Financial Analyzer — LangChain + Pinecone + OpenAI. Reduced research time from hours to minutes.
  • VRP 8-Algorithm Benchmark — Hybrid IP+GA identified as optimal.
  • Sanskrit Transformer — Built from scratch. Custom word2vec vs StanfordNLP + IndicNLP.
  • Alterlights Project Leader · Leetcode Group Member.
Ongoing
We Can We Will
Volunteer Teacher
  • English & Mathematics for 6th/7th grade students. Food drives and community outreach. The access problem at scale.
Capabilities

Skills with
Evidence

// Agentic AI Systems
LangGraphProduction
Built 3-agent lead nurturing system with shared state graph — qualification, technical, and pricing specialists. Architectural choice: LangGraph over LangChain because conversation state across 3 channels requires a graph, not a chain.
RAG Pipelines3× Built
Three production RAG systems: Qdrant + SentenceTransformers for conversational context, Pinecone + OpenAI for financial docs. Chunking strategy tuned per domain.
LangChainProficient
Used for linear document QA where LangGraph overhead wasn't warranted. Trade-off reasoning, not default.
// Computer Vision
DenseNet-121 + BayesianValidated
96.7% sensitivity, 97.1% AUC. Bayesian Gaussian Dropout adds uncertainty quantification — essential for clinical trust beyond raw accuracy.
Grad-CAM ExplainabilityClinical
Drove +63% radiologist acceptance. Explainability was the product — accuracy alone couldn't drive adoption.
MTCNN + FaceNetDeployed
Face detection + embedding pipeline for Attendance. SVM on embeddings + geolocation. 92% accuracy, live.
// NLP & LLM
Transformer ArchitectureResearch
Built from scratch for Sanskrit NLP — not fine-tuning. Architecture from primitives. Custom word2vec on web-scraped corpus.
OpenAI APIProduction
Production integration for Financial Analyzer with domain-specific system prompts and hallucination thresholds.
WhisperX + pyannoteReal‑time Audio
Built call scoring pipeline with speaker diarization and custom rubric‑based scoring. Redis for async job queue.
// ML Systems
Optimization AlgorithmsBenchmarked
8 algorithms for VRP with systematic methodology across runtime, quality, scalability. Hybrid IP+GA identified as optimal trade-off.
PyTorch / TensorFlowCore
Custom training loops, loss functions, gradient manipulation. Not just high-level APIs — comfortable at the primitive layer.
// Deployment & MLOps
Streamlit / GradioDeployed
Deployed Attendance and Pneumonia demos. Philosophy: deployed 92% beats 99% notebook.
Vector DatabasesProficient
Qdrant for conversational retrieval, Pinecone for document search. Selected per use-case — not one-size-fits-all.
RedisAsync Queues
Used in call scoring pipeline for job distribution and caching. Reliable and fast.
Talk to the work

Two AI Tools
Built for You05

Both tools run on Claude and are trained on my full project data — specific metrics, architectural decisions, and SARL case studies. The chat agent cites specifics, not summaries. The Fit Tool gives an honest match score with direct evidence from my work.

AI
Ask Sai's Agent
Trained on all 6 projects + internship systems · cites specific metrics and decisions
I'm an AI trained on Sai's project data. Ask about specific engineering decisions, trade-offs, or why he chose one approach over another. I'll cite real numbers.
Try asking:
FIT
Role Fit Assessment
Paste a JD → get honest match % with specific evidence
Job Description
Paste a JD above and analyze. You'll get an honest score with evidence from my specific projects.
Background

Education06

2021 — Present
B.Tech Artificial Intelligence
VijayBhoomi University, Mumbai
  • Alterlights Project Leader — student innovation initiative
  • Leetcode Group Member — algorithmic problem-solving
  • All major projects as independent research, not coursework
  • CV, NLP, multi-agent systems, optimization
Pre-University · KV No. 1, Mumbai
PCMCs — 85%
10th CBSE: 93.4%
  • Physics, Chemistry, Mathematics, Computer Science
  • Strong quantitative foundation before AI specialization
  • We Can We Will Foundation — volunteer English & Maths teacher
  • 6th & 7th grade. Community food drives.
Don't be a stranger
let's build something.

I love talking to people who care about what AI actually does in the world. AI/ML roles, agent consulting, research collabs — or honestly, just a great conversation about a hard problem worth solving.

📄 Download Full Resume
Email
ksaisovit@gmail.com
→
GitHub
github.com/ksaisovit
→
LinkedIn
linkedin.com/in/ksaisovit
→
Kaggle
kaggle.com/ksaisovit
→
© 2026 K Sai Sovit · AI / ML Engineer · Mumbai, India
Built with curiosity. Shipped with care.