AI Engineer · Pune, India · working remote with Seattle

I build AI agents that take actions, not just write answers.

Software engineer at SalesDuo. I design agent systems that read email, query databases, review code and close the loop on their own, and I ship products on top of them.

See the workAsk my agent
180min
Daily purchase-order workflow, down from 3 hours
PO agent · SalesDuo
0%
Faster internal answers across 100+ SOPs and docs
RAG knowledge system · SalesDuo
0%
Less pull-request review load on the team
AI PR reviewer · SalesDuo
1,500st
Place among 1,500+ hackathon participants
HireTainity

Showreel  /  20 seconds

01  /  Selected work

Products I designed, built and shipped.

Four live products. On the left of each, the real site captured from production. On the right, the system behind it.

01 / Live

SIEVE

Search your LLM can afford to use

The real site
How it works

A web search tool for LLM apps. It reads the pages, throws out the unreliable ones, cross-examines the rest and hands back one exact quote with its evidence.

  • Your model reads about 150 tokens, not 14,000
  • Says “nothing found” instead of guessing
  • Marks disputed answers openly
Claude Agent SDKSearchRAG tooling
▶ Product demoCase studyLive site

02 / Live

FlowLine

See what your agent will cost, and where it breaks, before you build it

The real site
How it works

A visual studio for agent workflows. Drag blocks for AI steps, tools, memory, approvals and quality checks onto a canvas, simulate the run, and see the cost before writing code.

  • Blocks for AI, tools, knowledge, flow control and checks
  • Human approval gates and AI-judge scoring
  • Simulate, import, export and share
Claude Agent SDKMCPCanvas
▶ Launch filmCase studyLive site

03 / Live

Clinic Studio

Clinic websites in 48 hours, sold by an agent

The real site
How it works

Ready-made websites for Indian doctors, live on their own domain in 48 hours. Behind it, an outreach agent finds doctors in a chosen city, collects their email and WhatsApp, and sends each one a tailored message.

  • Agent finds doctors city by city and contacts them
  • 10 templates with WhatsApp booking built in
  • Loads in under 2 seconds on 4G
Outreach agentWhatsAppLocal SEO
Case studyLive site

04 / Web live · app soon

TryOn Now

See it on you first

The real site
How it works

AI virtual try-on. Shoppers see how a product looks on them before they buy. The web version is live and the mobile app is on its way.

  • Three photos in, thirty outfits out
  • Built for shoppers who return what doesn't fit
  • Mobile app going live soon
Generative AIConsumerMobile
Case studyLive site

Interlude  /  Anatomy of an agent

Seven layers, every time.

Keep scrolling · the stack opens layer by layer

Every agent I build has the same seven layers.

02  /  Agent architectures

How I design systems that act.

Eight agent systems I designed, with their branches, loops and failure paths. Pick one, run a scenario through it, and click any step to see why it is there.

Reads incoming purchase-order emails and carries each one to a committed transaction and a reply, with nobody in the loop. Invalid orders and API failures each have their own path.

Production system3 hours of daily work → 0

13 nodes · 14 edges · pick a scenario

Run:

Swipe the diagram to see the full system →

not a POinvalidAPI errorTRIGGER
Cron schedule
TOOL
Email ingest
LLM STEP
Is this a PO?
RESULT
Leave for a human
LLM STEP
Identify customer
TOOL
Excel extraction
LLM STEP
Read PO details
CHECK
Validate PO
RESULT
Ask client to fix
TOOL
Commit via SP-API
CHECK
Retry with backoff
STORE
Audit log
RESULT
Reply to client

Run log · Valid order10 steps · 3 LLMs · 3 tools · 1 check

  1. Pick a scenario above to walk a request through the system.

Illustrative data

Inspect

Click any node to see what it does and why it is there. Dotted lines are loops and fallbacks.

Node.jsOpenAICronExcel parsersAmazon SP-API

03  /  Live agent

Don't take my word for it. Ask the agent.

This is a real model running a tool loop over the same data this page is built from. It can search my projects, open an architecture and pull my experience, and it shows each call.

taj-agent · tool loop · 4 tools

Ask anything about my work. You will see every tool call it makes.

›

Answers come only from tool results · rate limited · no conversation is stored

04  /  Lab

Experiments and notes.

Things I test in public: what small decision models can do, what they cost, and where they get it wrong.

Video7 Oct 2026

FlowLine launch film

78 seconds on FlowLine: describe an agent in plain words, watch simulated requests run through it, see what it will cost, and export to the framework you already use.

Watch and read

Video7 Oct 2026

SIEVE demo: three questions, three answer shapes

A walkthrough of SIEVE on a yes/no question, a how-to and a comparison, showing the sources it reads and the ones it throws out.

Watch and read

Experiment23 Sep 2026

FormPilot: a form filler that decides instead of generating

A Chrome extension that fills any web form from a profile saved on your device. A decision model picks which value goes in which field. Nothing is generated and nothing is submitted for you.

Watch and read

Experiment22 Sep 2026

Can a model pick Show HN winners from the title alone?

5,200 real Show HN posts, title only, $0.16 and under six minutes. It ranked them better than chance and was overconfident about how likely success is.

Watch and read

Video21 Sep 2026

Generate vs decide

An LLM and Jev end in the same softmax. One scores about 100,000 tokens, the other scores 3 labels. That split changes what each is good for.

Read it

05  /  Experience

Where I've worked.

  1. Jul 2025 – Present

    Remote · Seattle, USA

    SalesDuo · Software Engineer

    • Architected a fully autonomous purchase-order system: it reads the emails, identifies the customer, validates the PO, extracts the Excel data, commits through Amazon SP-API and replies to the client. A 3-hour daily workflow now takes no human time.
    • Built a production AI SQL agent with schema discovery, query planning, multi-hop reasoning, error correction and execution tracing, removing the dependency on manual data pulls.
    • Engineered a company-wide RAG knowledge system over 100+ SOPs, configs and docs with Pinecone hybrid search and caching. Internal query resolution time fell 65%.
    • Developed an autonomous AI pull-request reviewer that does diff analysis, complexity estimation, regression detection and secure-code checks. PR review load dropped 70%.
    Node.jsOpenAIPineconeMySQLAmazon SP-APIGitHub API

  2. Mar 2025 – May 2025

    Remote · Bengaluru

    CapitalXAI · Fullstack Engineer Intern

    • Built an end-to-end transcription and audio intelligence platform with real-time streaming, session persistence and secure cloud storage.
    • Implemented credit-based metering and subscription billing on Stripe webhooks, with automated reconciliation and event-driven state sync.
    • Led UI engineering for app.engagexai.com and built the Campaign Composer, which orchestrates multi-step async pipelines and polling workers.
    ReactSupabaseDeepgramStripeVercel

  3. Feb 2025 – Jun 2025

    Pune

    Accenture · Associate Software Engineer Intern

    • Automated enterprise SAP workflows for telecom operations: order creation, upgrades, suspensions and IMEI handling.
    • Wrote automation scripts in Java on enterprise testing frameworks, improving regression coverage across multi-module systems.
    JavaSAPTest automation

  4. May 2024 – Jun 2024

    Remote · Pune

    Tradlie LLP · Full-Stack Developer Intern

    • Built a production dashboard and analytics UI.
    • Optimised GraphQL queries through Apollo Client, improving data access latency by 40%.
    ReactTailwindRechartsGraphQL

Education

B.Tech, Information Technology

Pimpri Chinchwad College of Engineering, Pune

2021 – 2025 · CGPA 8.53 · top 5% of department

Recognition

  • 1st place among 1,500+ participants with HireTainity
  • 30+ projects built: agents, RAG systems, automation tools and AI full-stack apps
  • 98.3 percentile in MHT-CET
  • 500+ DSA problems solved across LeetCode, Codeforces and CodeChef
  • Web Dev Lead: mentored 100+ students and organised BlueBit 2.0
Me and my stack, levitating · made with Grok

Toolbox

AI / LLM
OpenAI · Claude Agent SDK · RAG · LangChain · Embeddings · Pinecone · MCP
Backend
Node.js · FastAPI · Express · Supabase · Firebase
Languages
TypeScript · JavaScript · Python · SQL · Java · C/C++
Data
PostgreSQL · MySQL · MongoDB
Infra
Docker · AWS EC2 · Vercel · Cron workers · Event pipelines
Frontend
React · Next.js · Tailwind

06  /  Archive

Everything I've built.

The current products first, then earlier work back to my first React apps. The pictures are the live sites as they look today.

25 projects

Early apps

Water Jug SolverFintrackSwiggatoExpense ManagerCalculator

07  /  Contact

Building something with agents? Let’s talk.

I’m open to AI engineering roles and to startups that want an agent that takes real actions and costs less to run.

tajmalnas36@gmail.com