Ayush Patel

AI/ML • Senior • Bengaluru, India • UTC+05:30 • Work permits: India • Available in 2 weeks

AI Engineer

5+ years exp

Python

Java

MLOps & Deployment: 5/10

Data Pipeline & Feature Engineering: 4/10

Active 11 days ago

Invite to interview

Download CVCV

Overview

Technical skills

Timeline

Roles

Overview

LLM-focused backend engineer (senior) experienced building agentic multi-node pipelines for production voice-first workflows; strongest at integrating hosted LLMs, observability, and external API fallbacks. The most proven skill is architecting and operating an async LangGraph-based multi-agent orchestration with concrete artifacts in agent_graph.py and a comprehensive e2e QA runner in e2e_test.py. There is little or no evidence of custom model training, research-grade experiments, or GPU/quantization engineering in public code.

Phone

Technical skills

Languages

Python

Java

Python

Pydantic

SQLAlchemy

Asyncio

HTTPX

Alembic

Uvicorn

AI/ML

LLM

Gemini

LangChain

RAG

Model Context Protocol

Prompt Engineering

LangSmith

NLP

Text-to-Speech

Amazon SageMaker

AWS Bedrock

Claude

CrewAI

LangGraph

Llama

OpenCV

PyTorch

TensorFlow

Transformers

Agentic Workflows

Multi-Agent Systems

Tool Use

PEFT

Qwen

DevOps

GCP

Google Cloud Run

AWS

Kubernetes

Other

PostgreSQL

SQLite

Anthropic

Hugging Face

Ollama

OpenAI

GitHub

Containers

WebSockets

Grafana

Databases

Apache Kafka

Weights & Biases

Context Engineering

Docker

Sentry

Groq

Prometheus

Redpanda

AI Agents

Loki

Deepgram

Fine-tuning

Google ADK

Multimodal AI

OCR

Sentiment Analysis

Timeline

AI Engineer • Middle

Cyble • Full-Time

Apr 2026 to Present 6 Months Bengaluru In office

Built LLM-driven alert processing flows combining deterministic tagging with LLM reasoning to surface relevant events and reduce operational noise. Worked on Kafka/Redpanda queue monitoring, including consumer lag and queue health, and supported reliability improvements in production. Managed Kubernetes deployments by tuning CPU/memory, scaling replicas, and debugging distributed issues using application logs and traffic visibility. Implemented multimodal vision workflows for brand/logo misuse detection with OCR and automated image preprocessing, improving observability and debugging speed.

Apache Kafka

Redpanda

Kubernetes

Prometheus

Grafana

Loki

Sentry

OCR

Model Context Protocol

Senior LLM Engineer • Senior

Infocusp Innovations • Full-Time

Nov 2024 to Apr 2026 1 Year 5 Months Pune In office

Developed agentic and multi-agent workflows focused on improving reasoning quality, output fidelity, and response latency. Implemented RAG and context engineering approaches to enhance accuracy while maintaining reliability in distributed environments. Built supporting components for code sandboxing, LLM evaluation, and observability, and incorporated MCP and tool-use patterns in production-oriented research.

Agentic Workflows

Multi-Agent Systems

RAG

Context Engineering

Model Context Protocol

Tool Use

Data Scientist • Middle

SquadStack • Full-Time

Sep 2023 to Oct 2024 1 Year 1 Month In office

Worked on LLM and agent quality evaluation, using quality and business metrics to identify improvement areas and reduce manual coordination effort. Built solutions involving RAG, prompt engineering, and NLP for knowledge and generation tasks. Contributed to voice-AI workflows including speech-to-text and text-to-speech components, supporting downstream NLP-driven applications.

RAG

Prompt Engineering

NLP

Text-to-Speech

LLM

Senior AI/ML Engineer Confidence: Medium LLM Engineer

Model Architecture & Training

2/10

How well models are designed and trained

Limited model engineering evidence; the code integrates hosted LLMs (Gemini) with structured output use but contains no model training, custom layers, or tuning pipelines.

Evidence

welfareflow-india/agent_graph.py: _generate_llm_reasoning — async Gemini Flash invocation with low-temperature settings

welfareflow-india/agent_graph.py: voice_intent_agent_node — llm_structured = _llm.with_structured_output(ExtractedCitizenProfile) and use of .ainvoke()

Data Pipeline & Feature Engineering

4/10

How data is prepared for models

Solid data-prep and domain-specific feature handling for names and documents, including custom phonetic normalisation and token-aligned similarity scoring.

Evidence

welfareflow-india/agent_graph.py: preprocess_indian_name — transliteration vowel/consonant normalisation and honorific stripping

welfareflow-india/agent_graph.py: _best_token_alignment_score and compute_name_match_score — token-aware Jaro-Winkler alignment used in document_audit_node

welfareflow-india/document_audit_node (agent_graph.py) — end-to-end OCR field mapping and cross-document name comparison

Experimentation & Evaluation

3/10

How results are measured and tested

Practical QA and reproducible acceptance testing for the pipeline exist, but no ML experiment tracking, ablation studies, or formal model evaluation harness is present.

Evidence

welfareflow-india/e2e_test.py: run_scenario / write_html_report — comprehensive end-to-end acceptance test scenarios and QA report

welfareflow-india/main.py: LangSmith integration and run_config metadata for pipeline runs (traceability at runtime)

MLOps & Deployment

5/10

How models are shipped to production

Strong production-oriented engineering: async FastAPI service patterns, background pipeline orchestration, observability/tracing hooks, DB session scoping, and multi-tier external API fallbacks.

Evidence

welfareflow-india/main.py: _lifespan, _run_pipeline, background tasks for SLA watchdog, and endpoint lifecycle handling

welfareflow-india/observability.py and agent_graph.py: LangSmith tracing (traceable decorators) and startup health checks

welfareflow-india/uipath_maestro.py: acquire_token token cache and tiered submission fallbacks

Computational Efficiency

2/10

How efficiently computing resources are used

Some asynchronous and efficiency-aware patterns (parallel OCR, retry/backoff, token caching) are used but there is no GPU/quantization/distributed training or low-level performance optimization.

Evidence

welfareflow-india/agent_graph.py: document_audit_node — parallel OCR via asyncio.create_task and asyncio.gather

welfareflow-india/agent_graph.py: _retry_async — exponential backoff wrapper used for external calls

welfareflow-india/uipath_maestro.py: in-memory token caching to avoid repeated auth round-trips

Research Depth & Innovation

3/10

Depth of research and new ideas

Some domain-driven algorithmic work (custom Jaro-Winkler and token-alignment strategy) shows thoughtful problem solving, but no novel research contributions or reproduced paper implementations are present.

Evidence

welfareflow-india/agent_graph.py: compute_jaro_winkler and _jaro_similarity — full Python implementation

welfareflow-india/agent_graph.py: _best_token_alignment_score — conservative assignment algorithm addressing token truncation and reordering

Expertise

AI Agents & Agentic Workflows• Senior

Industries

Government• Senior

Technologies

Databases

LangGraph

Weights & Biases

LangChain

Claude

Qwen

OpenCV

Groq

Model Context Protocol• since 2024

Loki• since 2026

Prometheus• since 2026

Fine-tuning

Prompt Engineering• since 2023

Multimodal AI

AI Agents

NLP• since 2023

LangSmith

Ollama

PEFT

Llama

AWS Bedrock

Transformers

TensorFlow

PyTorch

AWS

Docker

Kubernetes• since 2026

CrewAI

Gemini

Grafana• since 2026

LLM• since 2023

RAG• since 2023

Google ADK

Redpanda• since 2026

Sentiment Analysis

OpenAI

Anthropic

Hugging Face

Amazon SageMaker

GitHub

OCR• since 2026

Text-to-Speech• since 2023

Deepgram

Context Engineering• since 2025

Agentic Workflows• since 2024

Multi-Agent Systems• since 2024

Tool Use• since 2024

Deep Learning• mentioned only

LangGraph• mentioned only

Recommendations

Develop production-grade multi-agent LLM systems that require robust orchestration, tracing, and sandbox fallbacks (voice + OCR + RPA integrations).
Implement privacy-sensitive, compliance-first citizen-facing services that integrate external APIs and RPA (human-in-the-loop approval, consent revocation, masked PII vault).
Build integration layers and reliability logic for third-party services (retry/backoff, tiered fallbacks, token caching, health probes) and lead backend service design for LLM orchestration.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

patel-ayush welfareflow-india Face-Swap

Senior Data Scientist Confidence: Medium Data Engineer

Senior-level engineer specializing in production-grade async multi-agent pipelines for government welfare automation. The strongest proven skill is pipeline orchestration and integration, as evidenced by the LangGraph-based agent_graph.py together with the multi-tier UiPath integration in uipath_maestro.py. What is not evidenced is any statistical modeling or model-training work, formal hypothesis testing, or extensive unit-test coverage beyond the end-to-end QA runner.

Statistical Rigor

1/10

Correct use of statistics

Minimal statistical rigor evidenced; there are QA checks and scenario pass/fail assertions but no hypothesis testing, uncertainty quantification, or experimental design.

Evidence

welfareflow-india/e2e_test.py: scenario definitions and PASS/FAIL verdicts used for QA rather than statistical analysis

Data Wrangling & Cleaning

8/10

Preparing and cleaning data

Strong data wrangling and cleaning work for identity verification: a bespoke Indian name normaliser, a token-aware Jaro-Winkler alignment implementation, robust OCR integration and sandbox/mock fallbacks.

Evidence

welfareflow-india/agent_graph.py: preprocess_indian_name, compute_jaro_winkler, compute_name_match_score, _best_token_alignment_score

welfareflow-india/agent_graph.py: _call_sarvam_vision with mock OCR fallbacks and parallel OCR task orchestration

Exploratory Analysis & Visualization

4/10

Exploring and visualizing data

Exploratory outputs and human-oriented reporting are present but not research-grade EDA; there is a readable QA HTML report and a simple impact dashboard linking pipeline outcomes to business metrics.

Evidence

welfareflow-india/e2e_test.py: write_html_report generates a self-contained QA acceptance report

welfareflow-india/main.py: impact_dashboard aggregates case metrics and computes unlocked benefits

Predictive Modeling

1/10

Building models that predict

Uses LLMs for structured extraction and plain-language reasoning but shows no model development, feature engineering or predictive model evaluation; the ML work is orchestration and inference-only.

Evidence

welfareflow-india/agent_graph.py: _llm with Gemini Flash and llm_structured.ainvoke for ExtractedCitizenProfile schema

Business Insight & Impact

6/10

Turning analysis into business value

Clear product and policy-aware framing for a government welfare use case, with DPDP consent handling, human-in-the-loop gating, and impact-oriented metrics; analysis is tied to operational business outcomes rather than abstract modeling.

Evidence

welfareflow-india/main.py: initialize_case includes consent, OTP checks and DPDP-aware revoke_consent logic

welfareflow-india/agent_graph.py: eligibility_router_node implements PM-Kisan and Ayushman Bharat business rules and emits citizen-facing messages

Reproducibility & Notebook Hygiene

5/10

Clean, repeatable analysis

Reproducibility practices are moderate: pinned requirements.txt, sandbox modes and a runnable end-to-end QA runner; there is no explicit data versioning or CI shown but the system is designed to run in zero-infra sandbox mode.

Evidence

welfareflow-india/requirements.txt: pinned dependencies for reproducible environment

welfareflow-india/e2e_test.py: sandbox-mode scenarios and a runnable QA acceptance test suite

Expertise

Data Science• Middle

Streaming• Middle

Industries

Government• Senior

Farming & Agriculture• Middle

Financial Services• Middle

Health Care• Middle

Technologies

PostgreSQL

SQLite

Asyncio

HTTPX

Uvicorn

Alembic

Deep Learning• mentioned only

LangGraph• mentioned only

Recommendations

Develop real-time, event-driven backend services that integrate LLMs with external APIs and RPA systems, leveraging the existing LangGraph and UiPath integration patterns.
Build identity verification and name-normalisation modules for government or financial services, extending the token-aware Jaro-Winkler approach in agent_graph.py.
Implement human-in-the-loop workflows and compliance features for regulated domains, including consent management and auditable traceability.
Create robust end-to-end QA harnesses and sandboxed developer workflows, expanding the e2e_test.py patterns into CI jobs and reproducible test fixtures.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

patel-ayush welfareflow-india Face-Swap

Senior DevOps Engineer Confidence: Medium Generalist

Senior-level LLM and backend engineer focused on real-time audio integrations and recommendation engines with solid concurrency and test practices. The strongest proven skill is building a production-grade real-time Twilio-to-Gemini voice bridge with custom mulaw resampling and session management (voicebot-resume/app/server.py). Public code does not show infrastructure-as-code, CI/CD pipelines, cloud cost engineering, or documented runbooks and incident postmortems.

CI/CD Pipelines

Automated build and deploy

Not evidenced in public code

Infrastructure as Code

Managing servers with code

Not evidenced in public code

Containerization & Orchestration

2/10

Working with containers

Basic containerization evidence (single Dockerfile) is present but lacks multi-stage builds, non-root user, resource rationale, or orchestration manifests.

Evidence

voicebot-resume/Dockerfile: simple single-stage Python image and uvicorn run command

Observability & Monitoring

2/10

Watching system health

Lightweight observability is present via structured debug logging and NDJSON debug logging, but no SLOs, alerting rules, or dashboards as code were found.

Evidence

voicebot-resume/dev/poc.py:debug_log - NDJSON debug logger writing hypothesis entries

voicebot-resume/app/server.py:log - consistent operational log messages across async tasks

Reliability & Incident Response

4/10

Keeping systems up

Good reliability practices in application code: async task separation, background tool execution to avoid blocking, silence monitors, blip-handling and graceful shutdowns are implemented, but there are no runbooks, incident postmortems, or deployment rollback strategies in code.

Evidence

voicebot-resume/app/server.py:media_stream - offloaded tool execution (execute_and_respond) and monitor_silence with explicit thresholds and session cancellation

voicebot-resume/dev/poc.py:receive_audio and monitor_silence - blip counters, fatal error detection, and _handle_sigint for graceful shutdown

Cloud & Cost Optimization

1/10

Smart use of the cloud

Cloud usage appears in application-level code (Google APIs and a Cloud Run comment), but there is no evidence of autoscaling strategy, IAM least-privilege design, spot/eviction handling, or cost-measurement artifacts.

Evidence

voicebot-resume/app/server.py:book_appointment - uses google.auth.default and googleapiclient (Calendar API)

voicebot-resume/Dockerfile: comment referencing Cloud Run default port 8080

Expertise

Observability & Monitoring• Middle

Site Reliability Engineering• Middle

Industries

Artificial Intelligence• Middle

Technologies

Containers

Python• since 2021 • Senior

GCP

SQLAlchemy

WebSockets

Pydantic

Google Cloud Run

Recommendations

Develop real-time conversational voice agents and telephony integrations that require low-latency audio processing and robust session management.
Build recommendation services or ranking systems where the candidate can extend the existing weighted scoring engine and productionize it with database schemas and tests.
Harden production deployments by adding IaC (Terraform/CDK), CI/CD pipelines (parameterized reusable workflows), and secrets/workload-identity patterns for Google APIs.
Create observability runbooks: define SLOs, alerting rules, dashboards as code, and incident playbooks tied to the existing silence/connection error handling.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

patel-ayush voicebot-resume video-recommendation-system