Skip to content
Bangalore, IndiaAI & Agentic Systems Engineer

I’m Mehul Jariwala.

I build the infrastructure underneath AI products.

8+ years building retrieval that can be audited, agents that survive a restart, and systems people can depend on.

Mehul Jariwala
Built with teams at
RingCentralFasteroutcomesIBMPublicis SapientFitbots
15
AI systems, all runnable
824
tests, all green in CI
14/15
with zero runtime deps
0
API keys needed to run them

Selected work

I build AI systems.

All of them clone and run green with no API key — the offline providers are real implementations, not mocks.

0165 tests

Retrieval that grades its own results

agentic-rag-platform

Hybrid BM25 and dense retrieval fused by reciprocal rank, with a loop that re-queries when the retrieved context cannot answer the question. The published eval table shows hybrid losing to plain BM25 — left in, with the reason, rather than tuned until it flattered the architecture.

  • Python
  • FastAPI
  • evals
Source ↗
02110 tests

An OpenAI-compatible gateway, in Go

llm-gateway-router

Cost-estimated and latency-aware routing across providers, per-provider circuit breakers, response caching with near-duplicate matching, and per-tenant spend attribution. Zero dependencies, static binary in a scratch image, every test under the race detector.

  • Go
  • net/http
  • Prometheus
Source ↗
0367 tests

Agent runs that survive their own process

agent-orchestrator

A run is an append-only event log and state is a pure fold over it, so a run paused for human approval resumes in a different process entirely. CI asserts that property on every push.

  • Python
  • event sourcing
Source ↗
04130 tests

Eval gates that fail on regressions, not noise

llm-eval-harness

Wilson confidence intervals and McNemar's exact test, because on a 50-case suite a six-point accuracy drop is indistinguishable from a coin flip — and a gate that fails on noise gets deleted within a fortnight.

  • Python
  • statistics
Source ↗
0552 tests

Barge-in done the whole way

realtime-voice-agent

Full-duplex streaming speech to model to speech. Interrupting stops the audio, cancels generation, and truncates conversation history to what the caller actually heard — the third part has no symptom in a single-turn demo, which is why it ships broken.

  • Python
  • asyncio
  • VAD
Source ↗

Focused tools

PII detection that validates, not just matches

llm-guardrails

Luhn and Verhoeff checksums, so a 16-digit order reference is not reported as a credit card. Includes Aadhaar and PAN, which Western tooling omits and Indian product data is full of.

49 tests

Natural language to SQL, allowlisted

text2sql-agent

A statement is rejected unless it is recognisably a read. Repair is driven by the real database error, because a generic failure message just gets the same guess back.

51 tests

Prompts that cannot change under an experiment

prompt-registry

Content-addressed versions and sticky A/B assignment with no shared state, so a user does not get a different personality on their second message.

45 tests

Dying halfway never means paying twice

llm-batch-processor

fsynced checkpoints and exact resume across 100k-item batches, with a cost estimate before the run rather than an invoice after it.

27 tests

Quality, latency, index size and cost together

embedding-benchmark

Because a model two points better at four times the price is the wrong choice, and an index eight times larger is eight times the bill forever. Matryoshka truncation is measured, not assumed.

31 tests

Extraction you can audit

doc-intelligence

Every extracted field carries the character span it came from, so “why does the system think this invoice is £3,720” has an answer that is a line number, not a shrug.

36 tests

MCP servers, and a way to test them

mcp-server-toolkit

JSON-RPC from scratch against the 2026-07-28 spec, plus a 14-check conformance report covering what actually breaks interoperability — starting with notifications that must never be answered.

42 tests

Redelivery that does not re-bill the model

kafka-llm-pipeline

At-least-once delivery with effectively-once cost, via a content-keyed dedup store. Poison messages leave the partition instead of blocking everything behind them.

30 tests

SSE parsing that survives real networks

streaming-chat-ui

A stream yields arbitrary byte chunks, not messages. Tested one byte at a time, split across CRLF, and split through a multi-byte UTF-8 character — the cases that work on localhost and drop tokens in production.

53 tests

LoRA from its definition, verified

lora-toolkit

Gradients checked against finite differences and trained adapters measured against the Eckart–Young floor, landing within 0.0008 of the theoretical optimum.

36 tests

Shipped & in use

I ship open source.

react-doc-viewer

A React document viewer rendering PDF, DOCX, Excel, PowerPoint, Markdown, images, video, CSV and 20+ file types — with annotations, text search, dark mode and i18n.

19 on GitHubnpm ↗Source ↗

Beyond my own repositories

Small fixes.
Shared impact.

All contributions

Contributing to the tools other engineers build with, from React components and AI interfaces to monitoring and documentation.

7 selected merged pull requests · 6 upstream projects

Industry work

Built with teams.
Used in the real world.

FasterOutcomes · Legal technology

Legal AI platform

Document ingestion, citation-backed research, and draft generation connected through durable workflows and attorney review.

LangGraph / Temporal / Python / Elasticsearch

Read about the work

IBM Labs · Enterprise platforms

Customer analytics & observability

Event ingestion, predictive services, and interactive dashboards supported by metrics, logs, and distributed tracing.

Kafka / FastAPI / React / OpenTelemetry

Read about the work

Tesco · Through Publicis Sapient

Omnichannel grocery

Regional grocery experiences across the UK, Central Europe, and APAC, with delivery integrations and independently deployable frontends.

React / Micro frontends / Kafka / Webpack

Read about the work

Fitbots · Founding team

Employee OKR platform

Goal tracking, performance insights, and HR workflows supported by a reusable component library and accessible interfaces.

React / Next.js / Stencil / Tailwind

Read about the work

Experience

From first commits to enterprise platforms.

Both ends of the range: deciding what to build with nothing in place, and shipping inside systems that already carry real traffic.

  1. RingCentral

    Current

    Senior Agentic Systems Developer

    Agentic systems for a communications platform.

    2026 — present
  2. Fasteroutcomes

    Lead Engineer, GenAI · founding team

    Legal document ingestion, semantic search, drafting, and human review, orchestrated with LangGraph and Temporal.

    2025 — 2026
  3. IBM

    Senior Full-Stack / AI Engineer

    Customer analytics, event pipelines, and observability dashboards. Microservices, predictive features, and cross-functional engineering mentorship.

    2024 — 2025
  4. Publicis Sapient

    Senior Software Engineer

    Tesco grocery experiences across UK, CEU, and APAC: maps and delivery integrations, micro frontends, and enterprise identity tools.

    2022 — 2024
  5. Fitbots

    Founding Engineer

    Employee OKR SaaS: goal tracking, HR workflows, and a reusable component library built with accessibility in mind.

    2018 — 2022

Stack

Languages
PythonGoTypeScriptJava
AI & retrieval
RAGagentsevalsMCPLoRAvector searchLangGraphLangChainTemporal
Backend
FastAPISpring BootNode.jsKafkagRPC
Frontend
ReactNext.jsTailwind
Platform
DockerKubernetesTerraformAWSAzure
Observability
PrometheusGrafanaOpenTelemetrySplunkNew Relic
Testing
JestCypressPlaywrightPyTestSelenium
Data
PostgreSQLMongoDBElasticsearchPySpark

A little more about me

From the interface
to the infrastructure.

My work spans web products, AI workflows, and the platforms that keep them running. I bring that full-stack perspective to retrieval, automation, and developer tools.

AI & automation

Agents, retrieval, document intelligence, and workflows with human review.

Full-stack products

Accessible interfaces, reusable components, and APIs that connect products to their data.

Platforms & reliability

Cloud delivery, event pipelines, observability, and engineering mentorship.

Education

2017–2019

Master of Computer Applications

Jain University, Bangalore

Gold medalist · GPA 9.1/10

2014–2017

Bachelor of Computer Applications

Bhagwan Mahavir College, Surat

GPA 8.9/10

Recognition

Jain University

Academic gold medal

Recognition for academic achievement during my MCA.

IBM

Recognition & BluePoints Award

Appreciation for contributions to the team.

Tesco

Rewards and Recognition

Recognition for work on customer experiences.

Notes from building

I write, too.

All writing

Evaluations

How I Would Test an AI Agent Before Letting Customers Use It

Build a small, useful evaluation suite around outcomes, forbidden actions, and realistic failures.

Read article

Agent reliability

Your AI Agent Timed Out. Did It Still Send the Refund?

A practical design for agent retries, durable state, and duplicate prevention, using a customer refund workflow.

Read article

Context engineering

Your Support Agent Does Not Need the Entire Chat History

How to build a useful context packet for an AI support agent without losing customer commitments or exposing unrelated data.

Read article

Have something in mind?

We may work
well together.

AI infrastructure, agentic systems, and software built to last. Let’s talk about what you’re building.

Send me an email

Or reach me on LinkedIn