Available for new rolesSOFTWARE ENGINEER · AI & AUTOMATIONSOFTWARE ENGINEER · AI & AUTOMATION
IbuildAIthatworksoutside the demo.
I'm Naman, an AI & Automation Developer at IDeaS (a SAS company). I co-built a real-time AI voice sales agent, built Proofline to check every number an LLM writes, and shipped ClueCode, a subscription desktop AI app, solo.
Open toForward Deployed Engineer, Applied AI Engineer and Software Engineer roles.Forward Deployed Engineer rolesApplied AI Engineer rolesSoftware Engineer rolesForward Deployed Engineer roles
An outbound AI agent that phones leads, holds a scripted but adaptive qualification conversation, and hands qualified leads to a sales team. Streaming speech in and out, barge-in, and an LLM whose every action is validated before it happens.
~2.9 sSourceMedian from end of speech to the agent's first audio, 13 turns in the browser demo (Chrome, push-to-talk), Oct 2026. Per turn: STT 0.7–1.3 s, LLM 0.8–1.6 s, TTS first audio 0.8–1.0 s. median voice reply
121SourceAutomated tests passing at the commit that integrated FreJun (Oct 2026): telephony provider and routes, call lifecycle, engine regressions, prompt schemas. automated tests
Illustration of a live call between the AI agent and a lead.
weekly_summary.md4 ok2 wrong
Occupancy rose 5.2% to 66.8% vs last year, while ADR grew 3.8% year over year.
Room revenue reached $464.4K, 3.2% behind budget, and RevPAR improved vs last week.
real output · synthetic hotel data · repaired in one pass
Proofline output: numbers in an LLM-written sentence marked supported or wrong.
Checks every number an LLM writes into a business report against the data, explains how a wrong one is wrong, and repairs it. A CLI that can gate a pipeline, a Python library, and an MCP server agents can call.
98.9–100%SourceShare of injected numeric errors caught on three held-out phrasing families, first run each (github.com/NamanChordia07/Proofline, docs/BENCHMARK.md). errors caught, unseen phrasing
16–20%SourceRecall of a baseline that flags a number only if no fact anywhere in the data has that value. for a naive lookup
A Windows desktop AI assistant with subscriptions, designed and shipped solo: Electron client streaming multimodal Gemini answers, Next.js and Postgres backend, device-session leases, Razorpay billing.
270+SourceClueCode's recorded green run, Sep 2026: web 138, shared 13, desktop unit 67, Electron security and a11y 11, website E2E 43, installer and auto-update 7. automated tests
Report pipelines with LLM-written commentary, Playwright checks across Salesforce and a revenue-management platform, analytics tools, mailbox automation, and a legacy dashboard modernised to Angular and Spring Boot.
~3 hoursSourceManual tracking effort the service-delivery team no longer does, as estimated when the flows replaced it. manual tracking removed
PythonPlaywrightPower AutomateStreamlit
Read the case study
~2.9 sSourceMedian from end of speech to the agent's first audio, 13 turns in the browser demo (Chrome, push-to-talk), Oct 2026. Per turn: STT 0.7–1.3 s, LLM 0.8–1.6 s, TTS first audio 0.8–1.0 s.
median from a caller finishing a sentence to the voice agent's first audio
98.9–100%SourceShare of injected numeric errors caught on three held-out phrasing families, first run each (github.com/NamanChordia07/Proofline, docs/BENCHMARK.md).
of injected numeric errors Proofline caught on phrasing it had never seen
270+SourceClueCode's recorded green run, Sep 2026: web 138, shared 13, desktop unit 67, Electron security and a11y 11, website E2E 43, installer and auto-update 7.
automated tests behind ClueCode, a desktop AI app I shipped solo
~3 hoursSourceManual tracking effort the service-delivery team no longer does, as estimated when the flows replaced it.
of manual mailbox tracking removed at IDeaS
02EXPERIENCEEXPERIENCE
Where I've worked.
Security, full-stack and BI internships, then full-time on AI and automation at IDeaS, a SAS company.
Nov 2025 – Present
IDeaS Revenue Solutions
AI & Automation Developer
IDeaS Revenue Solutions (a SAS company) · Pune, India
Engineered a Dockerized Python pipeline of 4 modular services that generates client forecast-review reports in 3 formats (PDF, Word, Excel) with LLM-written commentary, delivered through SIT, UAT and hypercare.
Developed Playwright end-to-end checks that validate client setup (account numbers, integration types) across 3 systems: Salesforce, the G3 revenue-management UI and its database.
Created a Streamlit analytics app (pandas, Altair) of 4 services (config, validation, calculation, charts) for best-available-rate trends and pricing-decision validation.
Jul 2025 – Oct 2025
IDeaS Revenue Solutions
Business Intelligence & Automation Intern
IDeaS Revenue Solutions (a SAS company) · Pune, India
Streamlined a shared customer-care mailbox with scheduled Power Automate flows that log incoming emails to Excel and reconcile replies from Sent Items, saving the team about 3 hours of manual tracking.
Jan 2025 – Jun 2025
IDeaS Revenue Solutions
Software Developer Intern
IDeaS Revenue Solutions (a SAS company) · Pune, India
Migrated the legacy JSP/Struts “At-a-Glance” dashboard to Angular and Spring Boot: Angular components and REST APIs over optimized SQL and Java Streams.
Extended the APIs from single-day to custom date ranges and released the new UI behind a database-driven feature toggle for a staged rollout.
Jul 2024 – Dec 2024
IFM Engineering Pvt. Ltd.
Security Software Intern
IFM Engineering Pvt. Ltd. · Pune, India
Identified and mitigated 3 risk classes (XSS, SQL injection, misconfiguration) in internal security audits; automated vulnerability scanning.
2025
Education
Vishwakarma Institute of Information Technology, Pune
Technical Head, Cyber Cell, VIIT: ran 5+ security workshops and CTFs for 150+SourceParticipants across 5+ security workshops and CTF events run as Cyber Cell Technical Head. students.
Joint runner-up, Data Viz-a-thon (Power BI, Tableau).
03WHAT I DOWHAT I DO
Three kinds of problems I like.
The common thread: a workflow people depend on, a system that has to be right, and evidence that it is.
Applied AI you can trust
LLM features with the evaluation, guardrails and fallbacks that let them ship: grounded generation, verification, model fallback, prompt-injection defence.
Taking a manual, cross-system workflow and turning it into something scheduled, tested and boring: browser automation, report generation, mailbox and spreadsheet flows.
Full-stack systems that hold up: auth, payments, data models with real constraints, desktop security, and test suites that catch regressions before users do.
