Mohan Wakchaure Let's talk

AI & Automation Engineer

Hi, I'm Mohan. I build AI that asks first.

For 3+ years I've turned manual, compliance-sensitive work into governed LLM workflows, with guardrails, human review and audit trails. They run in daily production across 13+ teams.

View projects Download resume
Portrait of Mohan Wakchaure in a navy suit and white shirt, smiling, with green hills behind him

Shipped tools for teams at

Law Offices of Mary KennedyPrintoMeeshoExosphereX AIIIT Bombay

(01) About

I turn manual, compliance-sensitive work into AI workflows that people can trust.

I'm an AI and automation engineer who owns solutions end to end: requirements, build, deployment, monitoring and iteration. My work covers LLM tool calling and structured output, retrieval over vector stores, multi-provider routing with failover, and human approval steps.

My rule is simple. The model drafts, rules decide, and a person signs off. Everything leaves an audit trail.

  • 13+Teams using tools I shipped
  • 50+Mailboxes scanned by AI
  • 100K+SKU rows checked nightly
  • ~77%Cloud Run spend cut

(02) Stack

The tools I reach for, day to day.

  • Anthropic Claude
  • Groq / Llama
  • Gemini
  • Mistral
  • ChromaDB
  • RAG
  • Tool calling
  • Python
  • FastAPI
  • Flask
  • PostgreSQL
  • SQLAlchemy
  • Microsoft Graph
  • Google Cloud Run
  • Cloud Build
  • WhatsApp Business API
  • Supabase
  • Firebase
  • Jira
  • Power BI

(03) Experience

Where I've been building.

  • Operations Automation Specialist — Law Offices of Mary Kennedy

    Sole automation engineer for a US immigration law firm. I own AI tools from requirements to deployment, working directly with the CEO, the front-desk lead and previous developers.

    FastAPI · Claude · ChromaDB · Cloud Run

  • Data Analyst / Automation Engineer — Printo

    Sole AI and automation contact across 13 teams. I mapped manual processes with product, operations, procurement and finance, then shipped 6 internal tools into daily use.

    Python · ETL · Groq · WhatsApp API

  • Fulfillment Analyst — Meesho

    Automation contact for IT, Operations, Sales, HR and Infrastructure. I automated Jira incident workflows and Slack requests, and built SLA and compliance trackers.

    Jira · Slack · SQL · Power BI

  • Machine Learning Intern — ExosphereX AI

    Built and evaluated ML models for internal AI product features, and integrated the outputs into production workflows.

    Python · scikit-learn

  • Software Engineer — Skylex Technology

    Analysed campaign performance data and built automated reporting pipelines that informed spend-reallocation decisions.

    Python · Reporting

  • Machine Learning Intern — IIT Bombay

    Supported AI and LLM features in internal tools and research workflows, and built RPA workflows with Automation Anywhere.

    Python · RPA

(04) Selected work

A few systems I'm proud to have shipped.

All of them run in daily use. Client data is never shown on this page.

How a case moves through the system

  1. AI extracts facts from CV text
  2. Precedent retrieved with citations
  3. Rules engine scores the criterion
  4. A reviewer decides

Amber dot is AI, green dot is rules, blue is a person.

Legal tech

EB-1A case-evaluation platform

Law Offices of Mary Kennedy

An LLM was directly deciding legal eligibility, which is a compliance risk. I replaced that with a deterministic rules engine and mandatory human review, and limited the AI to extraction and advisory drafts.

FastAPI · Claude · PostgreSQL · ChromaDB · Cloud Run

How it's built
  • ChromaDB retrieval grounds every suggestion and every human decision in cited precedent.
  • Session auth, owner, reviewer and front-desk roles, race-safe task claiming, live scoring and full audit logging.
  • Bulk import of about 100 messy Excel leads with email and phone dedup and a mandatory dry-run preview before any write.
  • Cloud Run compute spend cut about 77%, verified against billing, with zero-downtime deploy and rollback.

What happens to each client email

  1. Scan 50+ Microsoft 365 mailboxes
  2. Rule-based filtering
  3. LLM summary, type and urgency
  4. Twice-daily digest to the front desk
  5. Supervisor opens or forwards the original

AI productivity

Email-intelligence system

Law Offices of Mary Kennedy

Summarizes and classifies incoming client email by type and urgency, then delivers a timezone-aware digest so the front-desk team sees what matters first.

Python · Flask · Microsoft Graph · Groq · Mistral · Gemini

How it's built
  • Multi-provider LLM pipeline: Groq with Llama as primary, Mistral and Gemini as failover.
  • Stateless, token-based action links let supervisors view or forward the original message with no login.
  • A time-bound email cache, running as a Linux systemd service.

What runs every night

  1. Python ETL reads 100,000+ SKU rows
  2. Threshold check finds stock at risk
  3. Alert on WhatsApp Business API
  4. Alert in Google Chat

Operations

Low-stock alert pipeline

Printo

Replaced a manual daily check by three people with a nightly pipeline and real-time alerts. No stockout alert has been missed since go-live.

Python · ETL · WhatsApp Business API · MySQL · Supabase

From PDF to spreadsheet

  1. Invoice or PO, digital or scanned
  2. pdfplumber or OCR reads the text
  3. LLM extracts structured fields
  4. Auto-match against delivery challan
  5. Export to Excel

Finance automation

LLM invoice and PO extraction

Printo, for Finance and AP

A Streamlit tool that pulls fields out of invoices and purchase orders and matches them to delivery challans, cutting manual data entry.

Streamlit · Groq · pdfplumber · pytesseract · openpyxl

(05) Also shipped

Smaller tools, same standards.

Missed-call tracker

Pulls and dedupes call records from a telephony API, reconciles them against a live sheet, and sends twice-daily prioritized callback digests. My audit tooling traced a recurring trigger failure to a Google Sheets row-group-depth limit.

Apps Script · Vitel API · Google Sheets

Timekeeping summaries

An LLM condenses daily task logs into completed, pending and carried-over summaries. It falls back to rules on any API error, and billable hours always come from database totals, never from model output.

FastAPI · SQLite · Cloud Run · Gemini

Approval routing and material classification

Human-in-the-loop approval email workflows, plus a Firebase-backed classification and inventory tool used by about 10 procurement and operations people.

Python · Firebase · Firestore

Jira and Slack automation

An auto-response bot, round-robin and shift-based ticket routing, and automated acknowledgements, plus SLA trackers the procurement review adopted.

Jira · Slack · SQL · Power BI

(06) Passion

When the laptop closes, I ride.

Open roads, rain clouds and a full tank. A few frames from the road.

Starts muted, because browsers block autoplay with sound. Press Sound to turn it on.

Close view of the motorcycle with its lights on as the rider walks up, carrying a helmet
A rider in a black hoodie and cap walking toward a parked motorcycle with its lights on, under a stormy sky
A motorcycle helmet with a Bluetooth headset on a cafe table beside a glass of masala chai
A rider in a full-face helmet and riding jacket taking a mirror selfie in an elevator

(07) Questions

Things people usually ask.

What kind of role are you looking for?

AI and automation engineer roles with a focus on GenAI and agentic systems. I'm based in Pune and open to remote work or relocation across India.

Do you let AI make decisions on its own?

Not in compliance-sensitive work. My pattern is that the model drafts, rules decide and a person signs off, with every step audited. In the EB-1A tool I replaced an AI-decides pipeline with a rules engine and mandatory human review.

Can you own a project end to end?

Yes. At the law firm and at Printo I was the sole automation engineer, covering requirements, build, deployment, monitoring and iteration, and working directly with the people who use the tools.

How do you keep AI systems reliable and affordable?

Every LLM call has a fallback: provider failover, or a deterministic rule-based path on any error. I deploy with zero-downtime rollouts and watch cost. Cutting Cloud Run compute spend by about 77% was verified against billing.

Have a manual process that follows a pattern?

Send it over. I'll tell you what can be automated, what still needs a person, and how to keep an audit trail either way.

Email me Download resume