Open to Software Engineer / AI-ML Engineer / Backend Engineer roles
$ whoami --software-engineer
Jayant Bhatia
Software Engineer
Specializing in AI / ML & Backend Systems
Software Engineer specializing in AI, backend systems, and intelligent applications. I design and build software end to end — APIs, data pipelines, and production services — with generative AI, RAG, and multi-agent systems as my strongest technical specialization. 2 years of industry experience across Primotech and Mindcode Labs, working backend APIs and mobile features alongside generative AI, RAG, and multi-agent systems.
// every system here ships with an evaluation harness, not just a demo
jayant@portfolio: ~/profile
$
>
>
>
>
>
>
$ ▍
// the shape of every project below
2
years of industry engineering experience
8+
end-to-end AI systems shipped
0.94
F1 on flagged-invoice fraud model
18+
public repositories on GitHub
3
MCP security implementations
3
compliance frameworks audited
Software engineer. AI is my specialization.
Three layers that make up the same job — not three separate things I do.
01
Software Engineering
Building reliable software — the foundation everything else sits on.
- Backend development
- REST API design
- System architecture
- Databases
- Testing & debugging
- Deployment
02
AI / Machine Learning
Intelligent capability layered on top of a working system, not in place of one.
- Generative AI & LLMs
- RAG & vector retrieval
- Multi-agent systems
- Applied ML & computer vision
- Model evaluation
- AI security & guardrails
03
Systems & Production
Designing for the version of the system that has to keep running.
- Docker & containerization
- Service integration
- Evaluation harnesses
- Audit logging
- Cloud deployment
- Config-driven environments
About
I'm a Software Engineer with a foundation in full-stack and backend development, and a specialization in applied AI — generative AI applications, retrieval-augmented generation, and multi-agent workflows that hold up outside a notebook.
Over the past 2 years at Primotech and Mindcode Labs, I've built backend REST APIs with FastAPI, shipped production RAG features grounded in vector search, and built LangChain/LangGraph pipelines that retrieve real documents instead of letting an LLM guess — moving between backend logic, data, and model work on the same features.
My starting discipline is software engineering — state, failures, APIs, deployment — and I carry that same discipline into how I build AI systems: grounded in real data, designed to be evaluated rather than just demoed, and built as part of a larger system rather than a standalone AI demo.
Currently focused on
- Backend engineering & API design (FastAPI, REST)
- Generative AI & LLM applications
- Retrieval-Augmented Generation (RAG)
- Multi-agent systems & tool orchestration
- Applied machine learning (classification, regression, CV)
- AI security, evaluation & guardrails
// how I build
01Ground everything
RAG over real documents and tool results, never an LLM guessing from memory.
02Evaluate, don't demo
Every system ships with an eval harness — retrieval accuracy, tool-selection, answer correctness, latency.
03Design for production
FastAPI services, Docker images, persisted artifacts — the same discipline as any shipped feature.
04Report the gap honestly
When a model overfits or a metric is weak, that diagnosis goes in the README, not under the rug.
05Security by design
MCP guardrails, prompt injection prevention, PII redaction, and audit logging are integral — not afterthoughts.
↻eval signal loops back into retrieval & prompting
Experience
2 years across two roles — backend APIs, mobile features, and applied AI on the same team, not in separate lanes.
Aug 2025 – Present
~1 year
AI Engineer
Primotech
- Built and integrated backend REST APIs with FastAPI, connecting mobile app features to production data services
- Built and deployed production-grade RAG features for mobile apps using LangChain and vector databases
- Implemented MCP guardrails and security layers for AI assistant tool execution
- Developed and maintained cross-platform app screens and features, working end to end with design and backend teams
- Wrote Python scripts to export and analyze crash logs, cutting bug-resolution time by 25%
Oct 2024 – Jul 2025
10 months
AI Engineer Trainee
Mindcode Labs Pvt Ltd
- Completed a structured training program building RAG-based chatbot systems with Python and LangChain
- Built and tested retrieval pipelines against vector databases to ground LLM responses in domain-specific documents
- Assisted with prompt engineering and response evaluation for accuracy and relevance, working alongside senior engineers on production-bound features
Education
B.Tech, Computer Science Engineering — CGPA 7.4
JMIT, Kurukshetra University
2021 – 2025
Certifications
Data Analytics Virtual Experience Program
Deloitte Australia (via Forage)
2025
Research Analyst Virtual Experience Program
JPMorgan Chase & Co.
2025
MCP Security Certification
Model Context Protocol
2025
AI Red Teaming Certification
AI Security Community
2025
From request to production
Not an AI-specific pipeline — the general lifecycle every feature on this site went through, with the AI/LLM layer sitting inside step 3, not replacing the other seven.
01
Requirements
Understand the actual problem, the constraints, and who the system needs to work for.
02
Architecture
Design services, APIs, data flow, and system boundaries before writing implementation code.
03
Implementation
Build the backend, data layer, and — where the problem calls for it — the AI/LLM components.
04
Integration
Connect databases, external services, and AI providers into one coherent system.
05
Evaluation
Validate correctness, retrieval accuracy, latency, and failure cases — not just the happy path.
06
Deployment
Containerize and ship, with config-driven environments instead of hardcoded assumptions.
07
Observability
Log, trace, and audit — so a broken system is visible before a user reports it.
08
Iteration
Feed real usage and evaluation results back into the next version of the system.
Projects
End-to-end systems spanning backend services, agentic LLM applications, applied ML on tabular data, and computer vision — each built with an API/data layer underneath the model, and shipped with an evaluation harness rather than just a demo.
flight_bot · agent_graph
PLANNER
classify user intent
↓
RETRIEVAL
policy RAG · embeddings + cross-encoder rerank
↓
AGENT
tool-calling · Groq / Llama 3.3 70B
↓
TOOLS
flight · hotel · cost · booking lookup
↓
VALIDATOR
guardrails · answer traced to source
Multi-Agent AI Flight Booking Assistant
Multi-agent travel assistant grounded in real data and company policy
A LangGraph-orchestrated assistant that answers flight, hotel, cost, and travel-policy questions in plain English. A planner agent classifies intent, a tool-calling agent decides which live data source to hit, and a validator checks every answer before it reaches the user — so responses are backed by tool results or a retrieved policy document, never invented from memory.
- ▸LangGraph state machine: plan -> retrieve -> agent -> tools -> validate -> respond
- ▸Policy-aware RAG over a company knowledge base using embeddings + cross-encoder reranking
- ▸Real tool integrations (flight search via AviationStack, hotel search, trip-cost, booking lookup) with deterministic mock fallback
- ▸FastAPI backend service layer with a Streamlit front end, containerized with Docker Compose
- ▸Built-in evaluation harness scoring retrieval accuracy, tool-selection accuracy, answer correctness, latency, and token usage
97%
answers traced to a tool result or policy doc — never invented
demimonde · memory_window
you book a flight to Delhi next week
companion Recalled: window seat preference, ~₹8k budget, morning departures…
memory layer → embeddings persisted across sessions
Demimonde — AI Companion App
An LLM companion with persistent memory, not just a chat window
An AI companion mobile app built around natural, context-aware conversation rather than single-turn Q&A. The core engineering problem was memory: a persistence and personalization layer built on vector embeddings lets the companion recall past conversations and user preferences across sessions, instead of resetting context every time the app opens.
- ▸LLM integration via LangChain for natural, context-aware conversation
- ▸Persistent memory layer using vector embeddings for cross-session recall and personalization
- ▸Real-time chat interface backed by Supabase/Firebase for auth, data storage, and service integration
- ▸Cross-platform front end so the same conversational core ships to iOS and Android from one codebase
memory
persistent vector-embedding recall across sessions
Private / closed-source
invoice_intel · eval_metrics
0.94
F1 · flagged
95%
recall
94.2%
R² · freight
Invoice Intelligence
Freight-cost prediction and invoice fraud detection on real vendor data
Two production-style ML pipelines built on vendor invoice and purchase-order data: one regresses expected freight cost so inflated charges stand out, the other classifies invoices as normal or flagged using dollar-mismatch and receiving-delay signals. The core lesson of the project is feature engineering — explicitly encoding the gap between invoice and purchase totals as its own column.
- ▸Engineering a single Dollar_Gap feature took flagged-invoice recall from 17% to 95% on the same dataset
- ▸Tuned Random Forest classifier reaches 0.94 F1 on the flagged class (stratified 80/20 split, 4,878 invoices)
- ▸Model comparison for freight-cost regression (linear, decision tree, random forest) — best model reaches 94.2% R-squared
- ▸Saved scaler + model artifacts so inference always matches training-time preprocessing
- ▸RandomizedSearchCV hyperparameter tuning optimized on F1 to handle class imbalance
0.94 F1
flagged-invoice recall lifted from 17% -> 95% by one feature
tb_model · summary
~60K params · trains on CPU in under a minute
Tuberculosis Detection from Chest X-Rays
A ~60K-parameter CNN that trains on a CPU in under a minute
A deliberately lightweight medical-imaging classifier, built as a counterpoint to reaching for a 50M-parameter pretrained network by default. Chest X-rays are preprocessed with OpenCV (CLAHE contrast enhancement) and classified by a small custom CNN, then served through a Streamlit app that runs locally or in Docker — an exercise in matching model size to the data you actually have.
- ▸Custom 3-block CNN with global average pooling — ~60K parameters, trains in seconds on CPU
- ▸0.87 test ROC-AUC / 0.89 test PR-AUC on a stratified, class-balanced 400-image dataset
- ▸OpenCV preprocessing pipeline (CLAHE, resize, normalization) plus flip/rotation/brightness augmentation
- ▸Weighted cross-entropy, early stopping, and best-checkpoint saving for a small, imbalanced-risk dataset
- ▸Exported as a single joblib artifact shared by the CLI predictor and the Streamlit UI
0.87
test ROC-AUC with a ~60K-parameter CNN
roofsense · model_comparison
identical 70/15/15 stratified splits
LightGBM
ROC-AUC 0.85
PyTorch MLP
ROC-AUC 0.88
~10-pt train/test gap diagnosed as label noise — not hidden
ROOFSENSE
Rooftop solar viability, from a classifier to a physics-informed yield estimate
A rooftop solar viability estimator that compares a gradient-boosted tree baseline (LightGBM) against a PyTorch MLP head-to-head, then feeds the predicted viability into an NREL PVWatts-style formula to turn a probability into an actual 'X sqm usable roof -> Y kWh/year' report. A real overfitting gap surfaced during evaluation and is diagnosed in the README rather than hidden.
- ▸LightGBM vs. PyTorch MLP compared on identical stratified 70/15/15 splits and metrics
- ▸Diagnosed a ~10-point train/test ROC-AUC gap to injected label noise rather than reporting a cherry-picked number
- ▸Physics-informed solar-yield formula (area x irradiance x panel efficiency x performance ratio) layered on top of the classifier
- ▸Config-driven training, fitted preprocessing reused at inference, and per-run experiment logs for comparing training attempts
- ▸Fully Dockerized workflow with a Streamlit UI and an inference-only demo notebook
A/B
LightGBM vs PyTorch MLP on identical 70/15/15 splits
flight_bot · agent_graph
PLANNER
classify user intent
↓
RETRIEVAL
policy RAG · embeddings + cross-encoder rerank
↓
AGENT
tool-calling · Groq / Llama 3.3 70B
↓
TOOLS
flight · hotel · cost · booking lookup
↓
VALIDATOR
guardrails · answer traced to source
MCP Guardrail Server
Secure MCP server with command allowlisting and path protection
MCP (Model Context Protocol) server that executes only pre-authorized commands. Provides AI assistants with limited command execution capabilities in high-security environments. Features allowlisted bash commands, restricted file path access, and command execution timeout.
- ▸Executes only commands included in the allowlist (git, ls, mkdir, cd, npm, npx, python)
- ▸Restricts file access to pre-authorized paths only via JSON path configuration
- ▸Command execution timeout functionality for safety
- ▸Configurable via --allowed-commands and --path-config flags
- ▸Can be used with Claude Desktop for secure AI assistant integration
100%
commands blocked outside allowlist — zero unauthorized access
demimonde · memory_window
you book a flight to Delhi next week
companion Recalled: window seat preference, ~₹8k budget, morning departures…
memory layer → embeddings persisted across sessions
Agentic Red Team Security
AI red teaming framework for LLM agent vulnerability discovery
Framework that analyzes AI agent source code to discover tools, roles, and guardrails, then generates tailored attack chains to find vulnerabilities. Supports white-box and black-box testing of agentic AI applications.
- ▸White-box code analysis to discover tools, roles, and guardrails
- ▸Generates adaptive attack chains across 141 attack categories
- ▸Multi-round execution with crescendo escalation
- ▸LLM-powered judging against configurable policies
- ▸Integrates with 15+ penetration testing tools (nuclei, httpx, ffuf, impacket)
141+
attack categories analyzed
tb_model · summary
~60K params · trains on CPU in under a minute
VisionFlow
Production Python framework for real-time video stream processing with AI
Event-driven Python framework for building scalable, real-time video AI applications. Provides a clean backend abstraction layer for video ingestion, AI model inference, and multi-channel event distribution with full async/await support. Production-ready with YOLO, OCR, and Kafka integration.
- ▸Event-driven backend architecture for scalable video AI applications
- ▸Async/await support for non-blocking video processing
- ▸Integration with YOLOv8, OCR, and Kafka event pipelines
- ▸Modular plugin system for AI model inference
- ▸Production deployment with Docker and Kubernetes
real-time
video frame processing with AI model inference
invoice_intel · eval_metrics
0.94
F1 · flagged
95%
recall
94.2%
R² · freight
LLM Security Gateway
Production-ready LLM security gateway with prompt injection and PII detection
Security gateway service that sits between users and AI models, inspecting prompts and responses for harmful content, PII, and unsafe outputs. Provides centralized backend control, logging, and governance for LLM API usage with multi-provider support.
- ▸Prompt injection defense with real-time scanning
- ▸Automatic PII detection and redaction
- ▸Sensitive content filtering for toxic/harmful outputs
- ▸Audit logging with latency and cost tracking
- ▸Multi-provider support (Groq, Ollama, OpenAI)
multi-provider
LLM gateway with unified security layer
roofsense · model_comparison
identical 70/15/15 stratified splits
LightGBM
ROC-AUC 0.85
PyTorch MLP
ROC-AUC 0.88
~10-pt train/test gap diagnosed as label noise — not hidden
NLP Emotion Intelligence
Multilingual sentiment and emotion analysis engine with 5+ languages
Production-grade multilingual sentiment & emotion analysis engine covering 5 languages (English, Hindi, Spanish, French, Japanese) plus Hindi-English code-switching. Built on XLM-RoBERTa with LoRA adapters, ONNX INT8 inference, and cross-lingual bias auditing.
- ▸Multi-task XLM-RoBERTa with dual sentiment + emotion heads
- ▸5+ languages plus Hindi-English code-switching support
- ▸ONNX INT8 quantization for low-latency inference (~3-4x smaller)
- ▸Cross-lingual fairness auditing with disparity tracking
- ▸Hyperparameter optimization with Optuna Bayesian search
5+
languages supported with code-switching
demimonde · memory_window
you book a flight to Delhi next week
companion Recalled: window seat preference, ~₹8k budget, morning departures…
memory layer → embeddings persisted across sessions
AI Inbox Agent
AI-powered inbox analyzer with classification, summarization, and reply suggestions
FastAPI-based productivity tool that classifies messages (Request, Informational, Reminder, Follow-Up, Urgent), generates one-line summaries, extracts action items, and suggests professional replies. Built with the OpenAI API on a FastAPI backend, with a vanilla JS frontend.
- ▸Smart message classification (5 categories)
- ▸One-line summaries for quick comprehension
- ▸Action item extraction from messages
- ▸Suggested reply drafting with natural language
- ▸Zero-setup productivity tool with localStorage persistence
5
message categories classified
flight_bot · agent_graph
PLANNER
classify user intent
↓
RETRIEVAL
policy RAG · embeddings + cross-encoder rerank
↓
AGENT
tool-calling · Groq / Llama 3.3 70B
↓
TOOLS
flight · hotel · cost · booking lookup
↓
VALIDATOR
guardrails · answer traced to source
Evidence Log Verification Engine
Tamper-evident audit log verification with hash-chained integrity
Backend framework for verifying AI audit logs with chain integrity, tamper detection, and independence checks. Implements EU AI Act Article 12 compliance for automatic event logging. Checks chain validity, content hash matching, and external anchor references for provable audit trail integrity.
- ▸Hash-chained audit log integrity verification
- ▸Tamper detection via content hash matching
- ▸Independence check with external anchor references
- ▸EU AI Act Article 12 compliance for event logging
- ▸Three-level verification: CHAIN, TAMPER, INDEPENDENCE
3-level
verification depth (chain, tamper, independence)
flight_bot · agent_graph
PLANNER
classify user intent
↓
RETRIEVAL
policy RAG · embeddings + cross-encoder rerank
↓
AGENT
tool-calling · Groq / Llama 3.3 70B
↓
TOOLS
flight · hotel · cost · booking lookup
↓
VALIDATOR
guardrails · answer traced to source
Agentic Compliance Copilot
AI-powered compliance assessment for SOC2, GDPR, HIPAA, ISO 27001, PCI DSS
Agentic AI compliance assessment platform that audits organization code and infrastructure for major compliance frameworks. Uses a multi-agent swarm backend architecture with specialized agents for planning, regulation RAG, extraction, gap analysis, and report generation. Produces audit-ready reports in minutes.
- ▸5 specialized AI agents working in parallel (planning, regulation RAG, extraction, gap analysis, report generation)
- ▸SOC2, GDPR, HIPAA, ISO 27001, PCI DSS compliance frameworks
- ▸Multi-agent swarm backend powered by Model Context Protocol (MCP)
- ▸Autonomous code & infrastructure audit
- ▸Audit-ready report generation in auditor-grade format
5
compliance frameworks assessed
Engineering for scale
Not a theoretical checklist — six concrete decisions pulled from the projects below, each about what happens after the demo works.
Evaluation over intuition
Retrieval accuracy, tool-selection accuracy, answer correctness, latency, and token usage are all measured — not assumed — before a system is called done.
Multi-Agent Flight Booking Assistant
Guardrails at the boundary
Answers that can't be traced to a tool result or a retrieved document are flagged rather than returned, and MCP tool execution is limited to an explicit allowlist.
MCP Guardrail Server
Cost & latency awareness
ONNX INT8 quantization and LoRA adapters keep an XLM-RoBERTa model fast enough for production inference instead of shipping the full-size model by default.
NLP Emotion Intelligence
Data integrity under audit
Hash-chained logs with tamper and independence checks, built to satisfy EU AI Act Article 12 event-logging requirements rather than an ad-hoc log file.
Evidence Log Verification Engine
Reproducible training, not one-off runs
Config-driven training, persisted preprocessing artifacts, and per-run experiment logs so a model result can be reproduced, not just remembered.
ROOFSENSE
Right-sized models
A ~60K-parameter CNN was chosen deliberately over a 50M-parameter pretrained network, matched to the size of the actual dataset available.
Tuberculosis Detection
Also on GitHub
more experiments, evals & pipelines worth a read
Skills
Software engineering and backend skills get equal billing here — AI/ML is the specialization, not the whole list.
Software Engineering
Backend & APIs
AI / ML & Agents
Applied ML & Data
AI Security & Compliance
Infrastructure & Cloud
Also comfortable with: Flutter · Dart · Mobile App Development — from an earlier full-stack & mobile background.
Building software systems that ship. Let's talk.
Open to Software Engineer / AI-ML Engineer / Backend Engineer roles. I usually respond within a day.
+91-8571857081