Hire Me →
Open to AI/ML Roles AI/ML Engineer Mumbai, India

I build AI systems that ship.

I get genuinely excited about AI that closes real gaps — not demos, not dashboards. Real systems. Real humans. Real stakes. A portfolio of production systems and research 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 · LLMs · Agentic AI
LocationMumbai, India · Remote OK
CurrentRetail Sync — AI/ML Engineering Intern (May 2026–Present)
Core StackPython · FastAPI · LangGraph · RAG · PostgreSQL · Redis · Qdrant · RabbitMQ · Docker
StatusOpen — AI/ML, RAG, Agentic AI & Automation roles
Contactksaisovit@gmail.com
// Three things I'm proud of
01Intelligence OS — a modular information-management platform with provenance-aware knowledge modeling.
40%Ranking precision improvement in Talent Rediscovery Engine 2.0 over keyword-only search.
95%Sales-agent workload reduction reported for the DSJ Keep Learning lead-nurturing workflow.
Say hello → ksaisovit@gmail.com
40% Better Ranking Precision 500 Candidates Ranked ~15s Ranking Time 95% Sales Workload Reduction 100/hr Calls Processed May 2026 Retail Sync · Current 40% Better Ranking Precision 500 Candidates Ranked ~15s Ranking Time 95% Sales Workload Reduction 100/hr Calls Processed May 2026 Retail Sync · Current
Deep Dive

Two Builds.
Closer Look.★

Intelligence OS — Information Management & Knowledge Graph Platform
FastAPI RabbitMQ SQLite LLM APIs Obsidian Provenance
Problem
External information is heterogeneous, duplicated and continuously changing. The system needs to preserve what was observed, where it came from and how entities and claims connect rather than only storing summaries.
Architecture
Source ingestion → RabbitMQ → normalization → deterministic deduplication → NLP extraction → LLM enrichment → entity/relationship resolution → knowledge projection. The pipeline is modular so individual stages can be tested and replaced without redesigning the whole system.
Data Model
Canonical records cover information items, entities, relationships, claims and observations. Source provenance and confidence are attached so downstream analysis can trace back to evidence.
Output
Structured knowledge is projected into Obsidian notes and wikilinks, giving the system a human-readable knowledge layer alongside its machine-oriented storage model.
Status
Currently being hardened with Docker-based CI/CD, Prometheus/Grafana observability and cloud deployment for pipeline health, throughput, latency and resource monitoring.
Lesson
Trust comes from structure. Deterministic processing and provenance make the LLM layer easier to evaluate and the resulting knowledge easier to audit.
Talent Rediscovery Engine 2.0 — Semantic Candidate Re-Engagement
Sentence Transformers Claude FastAPI RabbitMQ Streamlit RAGAS · Planned
Problem
Recruiting teams already have candidate data, but keyword-only search can miss relevant historical candidates. Rediscovery needs to combine semantic similarity with explicit skills, experience and location signals.
Ranking
Candidate matching combines semantic similarity, skills, experience and location rather than treating lexical overlap as the full relevance signal.
Outcome
40% improvement in ranking precision over keyword-only search, while ranking approximately 500 candidates in ~15 seconds.
Generation
Claude produces match explanations and personalized WhatsApp/email outreach. The reported outreach workflow produced a 3× higher candidate response rate.
Async
RabbitMQ handles asynchronous processing so larger candidate sets and outreach workflows do not block the application layer.
Next
Authentication, Docker/CI/CD cloud deployment, RAGAS evaluation and Prometheus/Grafana observability are being added as the system moves toward a more complete deployment.
Selected Work

Systems, Agents
and Research02

01
Flagship BuildInformation SystemsKnowledge Graph
Intelligence OS
FastAPI · RabbitMQ · SQLite · LLM APIs · Obsidian
An information-management and intelligence platform designed to turn noisy external inputs into structured, traceable knowledge. The system ingests RSS, YouTube, Reddit and web sources asynchronously, then moves information through deterministic, NLP and LLM stages for deduplication, entity extraction and relationship discovery.
ProblemRaw information accumulates faster than a researcher can read, connect, verify and organize it. A useful system needs both collection throughput and traceability.
PipelineRSS / YouTube / Reddit / Web → asynchronous RabbitMQ workflows → normalization → deterministic deduplication → NLP extraction → LLM enrichment → entity / relationship resolution → Obsidian projection.
Data ModelCanonical information items, entities, relationships, claims and observations, with source provenance and confidence scoring so generated intelligence remains attributable to an underlying source.
StatusActively being hardened for Docker-based CI/CD deployment with Prometheus/Grafana monitoring for pipeline health, throughput, latency and resource usage.
LessonAn intelligence system is not just a scraper + LLM. Canonical entities, provenance and deterministic stages are what make downstream reasoning trustworthy.
PythonFastAPIRabbitMQSQLiteObsidian
02
Production-OrientedRecruiting AI
Talent Rediscovery Engine 2.0
Semantic Matching · Claude · RabbitMQ · Streamlit
A candidate rediscovery system that combines semantic similarity with explicit skill, experience and location matching instead of relying on keyword search alone. It also generates evidence-backed match explanations and personalized outreach for re-engagement workflows.
ProblemPreviously-seen candidates are valuable but easy to miss when recruiters depend on lexical search. The ranking problem needs both semantic relevance and structured constraints.
RankingSemantic similarity is combined with skills, experience and location signals to produce a more useful candidate ordering than keyword-only retrieval.
Outcome40% improvement in ranking precision over keyword-only search; approximately 500 candidates ranked in ~15 seconds.
GenerationClaude generates match explanations and personalized WhatsApp/email outreach; reported response rate was 3× higher for the generated outreach workflow.
NextAuthentication, Docker/CI/CD cloud deployment, RAGAS evaluation and Prometheus/Grafana observability are being added as the system moves toward a more complete deployment.
Sentence TransformersClaudePandasRabbitMQFastAPI
03
CurrentText-to-SQL
Natural-Language POS Analytics
Retail Sync · FastAPI · PostgreSQL · pgvector / ChromaDB
A conversational analytics microservice for multi-store retail data. The system combines RAG, Text-to-SQL, schema linking and conversational context so users can ask natural-language questions about transactions, inventory and employee shifts without writing SQL directly.
ProblemOperational data is distributed across a complex relational schema. Users need answers in business language, but direct database access is difficult to use and hard to secure.
ApproachBuilt a pipeline around intent parsing, schema-aware SQL generation, retrieval and conversational context, backed by PostgreSQL and vector retrieval.
SafetyGenerated SQL is constrained toward secure, read-only analytics rather than arbitrary database operations.
InterfaceFastAPI backend with REST APIs / WebSockets and a Streamlit/React interface. Docker and OpenAPI are part of the delivery stack.
StatusEvaluation and performance metrics are being established as the service moves toward deployment.
FastAPIPostgreSQLpgvectorChromaDBText-to-SQLDocker
04
ShippedMulti-AgentRAG
Lead Nurturing Agent
DSJ Keep Learning · LangGraph · Qdrant
A multi-agent sales automation workflow coordinating lead qualification across WhatsApp, Email and SMS. The system uses LangGraph to manage stateful, branching conversations and RAG to retrieve context for personalized responses.
ArchitectureA 3-agent LangGraph workflow separates qualification, technical and pricing responsibilities while maintaining shared conversation state.
RetrievalSentence Transformers + Qdrant provide semantic retrieval for personalized responses.
ChannelsThe workflow coordinates WhatsApp, Email and SMS interactions rather than treating each channel as an isolated automation.
Impact95% reduction in sales-agent workload, 50–55% efficiency improvement and 20% increase in customer satisfaction were reported for the system.
LessonStateful orchestration matters when conversations branch, return to earlier decisions and span multiple communication channels.
LangGraphQdrantSentence TransformersPython
06
Document AIRAG
LLM Financial Analyzer
LangChain · Pinecone · OpenAI
A PDF-focused financial research assistant using retrieval-augmented generation and document-structure-aware chunking. The system was built around the idea that financial reports should be retrieved according to their structure rather than treated as generic text.
ProblemAnalysts spend substantial time navigating long annual reports to extract structured insights and compare information.
RetrievalPinecone stores embeddings while LangChain orchestrates retrieval and generation against financial PDFs.
ChunkingThe system uses domain-specific chunking designed around financial document structure rather than one generic chunk size.
OutcomeResearch workflows were reduced from hours to minutes in the project evaluation.
LessonFor changing knowledge bases, retrieval design and chunking can provide a more maintainable path than fine-tuning the model itself.
LangChainPineconeOpenAIRAG
07
Optimization Research
VRP — Eight-Algorithm Benchmark
GA · SA · ACO · Dijkstra · Hybrid Search
A systematic benchmark of eight approaches to the Vehicle Routing Problem, including individual metaheuristics and hybrid strategies. Instead of assuming one algorithm would dominate, the work compared trade-offs empirically across solution quality, runtime and scalability.
MethodsGenetic Algorithm, Simulated Annealing, Ant Colony Optimization and Dijkstra, plus four hybrid combinations.
BenchmarkApproaches were compared systematically rather than selected from theory alone.
FindingHybrid IP + GA emerged as the strongest overall trade-off in the benchmark.
LessonOptimization work often benefits more from a repeatable evaluation methodology than from arguing for a single theoretically preferred algorithm.
PythonDEAPOR-Tools
08
NLP ResearchBuilt From Scratch
Sanskrit Transformer
Custom Word2Vec · Transformer Primitives · Low-Resource NLP
A low-resource NLP project where the language model stack was constructed from primitives instead of relying on a pretrained transformer. The work also included a custom Word2Vec pipeline over a web-scraped corpus and comparisons against StanfordNLP and IndicNLP.
FoundationBuilt the Transformer architecture from scratch rather than fine-tuning a pretrained model.
DataCreated a custom Word2Vec pipeline over a web-scraped corpus for a low-resource language setting.
ComparisonCompared the approach against StanfordNLP + IndicNLP baselines.
LessonLow-resource NLP is often constrained by corpus strategy and data acquisition as much as model architecture.
PyTorchWord2VecTransformersBeautifulSoup
Recent Internship Projects

What I've been
building at work★

Natural-Language POS Analytics
Retail Sync · Current
Engineered a conversational analytics microservice that combines RAG, Text-to-SQL, schema linking and conversational context for multi-store transaction, inventory and employee-shift analytics. Built the FastAPI backend and Streamlit/React interface with REST APIs, WebSockets, Docker and OpenAPI; secure read-only SQL generation is part of the pipeline.
FastAPIPostgreSQLpgvector / ChromaDBText-to-SQLDocker
Multi-Agent Lead Nurturing
DSJ Keep Learning · Shipped
Automated lead qualification and nurturing across WhatsApp, Email and SMS with a LangGraph multi-agent workflow. The system uses specialist agents for qualification, technical and pricing conversations plus Sentence Transformers + Qdrant retrieval for personalized responses.
LangGraphQdrantSentence TransformersRAG
Call Intelligence Pipeline
DSJ Keep Learning · Async
Built an asynchronous call-processing pipeline using WhisperX for transcription, pyannote.audio for speaker diarization and Redis for job distribution. Concurrent processing supported approximately 100 calls/hour, with rubric-based scoring for agent evaluation.
WhisperXpyannote.audioRedisPython
CRM Query Assistant
DSJ Keep Learning · Secure RAG
Developed a RAG-powered CRM chatbot with PostgreSQL query validation, designed to answer operational questions while constraining database access to safe, read-only analytics. Automated evaluation shortened feedback turnaround from days to minutes.
RAGOpenAIPostgreSQLLangChain
Who I Am

Hi — I'm Sai.
I build for the real world.03

I'm an AI/ML Engineer focused on production-oriented LLM, RAG, agentic AI and automation systems. My work sits at the point where models meet software engineering: APIs, retrieval, orchestration, databases, asynchronous workflows, evaluation and deployment.

"The model is only one part of the system. The real work is making everything around it reliable, useful and easy to trust."

At Retail Sync, I'm currently building a natural-language POS analytics microservice that combines RAG, Text-to-SQL, schema linking, PostgreSQL and conversational context for multi-store business analytics. Before that, at DSJ Keep Learning, I worked on multi-agent lead nurturing, semantic lead scoring, call intelligence and a secure CRM query assistant.

Outside internships, my strongest independent builds are Intelligence OS and Talent Rediscovery Engine 2.0 — projects where I care as much about architecture and evaluation as I do about the model itself. I like building systems that can be explained, tested, observed and eventually deployed.

I'm looking for teams that care about shipping useful AI, are honest about engineering trade-offs, and want AI systems that survive contact with real users and real data.

Sai
K Sai Sovit
AI / ML Engineer · Retail Sync
Emailksaisovit@gmail.com
LocationMumbai, India
GitHubksaisovit
LinkedInksaisovit
StatusOpen to hire · AI/ML, RAG & Agentic AI
// How I approach engineering
01
Systems over demos — the useful part is everything around the model.
02
Architecture before tools — choose components for the problem, not the trend.
03
Evidence over intuition — evaluate retrieval, ranking, latency and failure modes.
04
Shipping is the final feature — but only when the system helps someone in practice.
Timeline

Where I've
Been Building04

May 2026 — Present
Retail Sync
AI / ML Engineering Intern
  • Natural-Language POS Analytics — RAG + Text-to-SQL + schema linking for multi-store transaction, inventory and employee-shift analytics.
  • Secure Query Pipeline — intent parsing, schema-aware SQL generation, pgvector/ChromaDB retrieval and conversational context with read-only constraints.
  • Application Layer — FastAPI backend and Streamlit/React interface using REST APIs, WebSockets, Docker and OpenAPI.
  • Deployment Readiness — evaluation and performance metrics are being established as the service moves toward deployment.
May 2025 — September 2025
DSJ Keep Learning
AI / ML Engineering Intern
  • Lead Nurturing — LangGraph multi-agent automation across WhatsApp, Email and SMS; reported 95% reduction in sales-agent workload.
  • RAG + Lead Scoring — Sentence Transformers + Qdrant personalization and asynchronous scoring with engagement metrics and exponential decay; 70% manual-effort reduction.
  • Call Intelligence — WhisperX + pyannote.audio + Redis concurrent processing, handling approximately 100 calls/hour.
  • CRM Query Assistant — secure RAG chatbot with PostgreSQL query validation and automated evaluation workflows.
2021 — Present
VijayBhoomi University
B.Tech. in Artificial Intelligence
  • Independent AI/ML Work — projects across LLMs, RAG, computer vision, NLP and optimization.
  • Intelligence OS — information management platform with asynchronous ingestion, provenance-aware knowledge modeling and Obsidian projection.
  • Talent Rediscovery Engine 2.0 — semantic candidate ranking, LLM explanations and automated re-engagement workflows.
  • Campus & Community — Alterlights Project Leader and Leetcode Group Member.
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

AI Tools
Coming Soon05

Coming Soon

Two interactive tools are in development: Ask Sai's Agent for project-aware Q&A and Role Fit Assessment for evidence-based job matching. I’m polishing the knowledge base, evaluation and deployment before exposing them publicly.

AI
Ask Sai's Agent
Project-aware Q&A with architecture, metrics and technical decisions
Coming Soon
Ask about the systems behind the portfolio.
The agent will answer questions about projects, trade-offs, stack choices and implementation details using a grounded project knowledge base.
Knowledge base · Evaluation · Deployment
FIT
Role Fit Assessment
Paste a JD → get an evidence-backed match assessment
Coming Soon
See where a role matches my actual experience.
The tool will compare a job description against specific projects, systems and skills — including strengths and honest gaps.
JD parsing · Evidence matching · Evaluation
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.