Gaurab DasAI Software Engineer

A more personal introduction

My work. My thinking.
A conversation away.

I work on enterprise software at Oracle and build practical AI projects you can explore, inspect and ask my AI twin about.

My workAnswers from my reviewed profileAn AI introduction, not a live conversation

A guided look at my experience

Gaurab’s AI twin - not a live conversation with Gaurab.

More profile topics

What would you like to know?

Ask about my experience, explore an AI project, or discover how I approach engineering decisions.

Profile topics are instant. Custom questions use AI and may take a little longer.

0 / 2,000

Custom questions are sent to OpenRouter and its selected model provider. Please avoid confidential information. Privacy details

Professional work

Product development and internal AI initiatives at Oracle.
Measurements and scope, with the context intact.

Reliable enterprise software

6 features and 150+ full-time defects resolved in Primavera P6 Professional.Read the context

I’m an Application Software Engineer 2 at Oracle in Hyderabad, working on Primavera P6 Professional since August 2023. I have 3+ years of experience building and deploying generative AI applications used by others. I delivered 6 features and resolved 150+ defects during full-time employment, separate from my internship contributions.

I strengthened import-type and user-privilege authorization, enhanced imports with project information such as Responsible Manager, and contributed to Delphi-to-C# modernization. I raised recurring average build success from about 50% to 90% by addressing build-machine failures; I also fixed Build Verification Test failures and authored FitNesse regression cases.

I work directly with customers to gather requirements, explain technical solutions, demonstrate features, and discuss troubleshooting and workarounds on calls.

Large datasets. Less waiting.

Delivered multiple performance improvements.Read the context

My most recent performance improvement addressed a large volume of data in a customer production environment. The shipped change reduced first project opening from approximately 52 to 16 seconds and project closing from 29-35 to 2-3 seconds. These are observed before/after measurements for that environment, and the customer expressed satisfaction with the improvement. I achieved these gains through query batching, deduplication, event-invalidated caching and deferred loading.

AI for the engineering team

BugTrace and 4 Codex plugins, used by 7 developers and 5 QA engineers.Read the context

I built and deployed BugTrace, an internal VM-hosted RAG tool that retrieves evidence from 5,300+ historical Primavera P6 bugs to support developer investigations. I developed the concept and integration with Codex assistance. Its documented retrieval uses OpenAI text-embedding-3-large through OCI, Chroma and SQLite FTS5. Developer review remains central; this internal prototype retrieves evidence rather than fixing bugs autonomously. It is not a publicly qualified production AI service. Confidential information in bug comments was redacted before creating vector embeddings.

I created an internal Codex marketplace with four plugins: BugDB authentication recovery, MCP-based TFS integration, IPS release automation, and P6 FitNesse test authoring using existing cases and QA fixtures. BugTrace and the marketplace are used by a shared group of 7 developers and 5 QA engineers, counted once across both tools. I built BugTrace and the four plugins independently, with AI assistance.

I collaborate with my team on the P6 AI chatbot, which is in internal beta. I contributed to its basic building blocks and am developing local HTML reports from LLM-generated instructions. The intended workflow keeps report rows out of model requests, with the aim of reducing token usage and cost. This reporting work is in development; cost savings have not yet been measured.

Build. Measure. Improve.

Independent projects built with AI assistance.
Public demos may sleep or reach free-provider limits.

PriceLab

Compare a small language model before and after fine-tuning.

  • QLoRA
  • PyTorch
  • Transformers
  • Gradio
Try PriceLab ↗
Architecture, evidence & limitations

PriceLab is my AI-assisted historical USD product-price experiment using Qwen3-0.6B, PyTorch, Transformers, Datasets, PEFT, TRL, bitsandbytes and Gradio. I prepared/audited data, ran guided Colab QLoRA training, recovered interrupted evaluation, compared baselines and selected the improved adapter. Retained splits: 19,991 training, 999 validation, 1,000 test examples. Initial training used one epoch; improved training used two. An epoch is one pass. QLoRA trains LoRA updates on a frozen 4-bit base.

Training checkpoints used validation loss; final candidate selection used validation price MAE. Validation MAE fell from $84.12 to $74.27 (11.71%); recorded GPU-float16 test MAE fell from $80.08 to $68.05 (15.03%). The test set was previously inspected. Only 24% were within 20%; estimates repeat $100 (298) and $120 (195). Ridge has slightly lower RMSE ($118.08 vs $118.97).

Public CPU website calls a private ZeroGPU worker with server-side owner authentication. Serving compares base with adapter off/on; it does not serve the merged export. A separately published standalone model merges the improved adapter into base weights without further training. Visitor use is bounded (10/session/hour, 100/day); free quota and sleeping hosts limit availability.

Open PriceLab demo ↗

I published a standalone merged fine-tuned Qwen3-0.6B model. It packages the improved LoRA updates into float32 base weights; merging is not new training. Local packaging compatibility checks compared 104 prompts. The website continues to use base-plus-adapter serving. Recorded benchmark metrics belong to the GPU-float16 adapter evaluation, not a new full benchmark of this export.

View published model ↗

IssueLens

Find the history behind a software issue, with evidence you can inspect.

  • RAG
  • LangChain
  • MCP
  • Chroma
Try IssueLens ↗
Architecture, evidence & limitations

IssueLens is my AI-assisted independent bug-evidence RAG application with 200 fictional issues. It uses Python, FastAPI, React/TypeScript, MiniLM embeddings, Chroma, SQLite FTS5 and reciprocal-rank fusion, with LangChain integrations, LiteLLM and metadata-only Langfuse. Keyword and semantic retrieval provide complementary evidence; citation validation does not prove answer correctness.

It is public on HF CPU Basic. Its separate local read-only MCP stdio server supports keyword/vector/hybrid search and issue lookup; it is not a hosted MCP endpoint. Two anonymous live assessments and the MCP modes passed dated checks. It contains no enterprise or confidential bug data.

Open IssueLens demo ↗

SupportDesk

A support agent that proposes actions. You decide what runs.

  • LangGraph
  • PostgreSQL
  • FastAPI
  • React
Try SupportDesk ↗
Architecture, evidence & limitations

SupportDesk is my AI-assisted independent customer-service demo using Python, FastAPI, React/TypeScript, LangGraph, Psycopg, PostgreSQL 18, Pydantic, LiteLLM and metadata-only Langfuse. Backend policy checks, explicit approval of the exact proposal, transactions and idempotency protect simulated returns/replacements. Ownership checks blocked unauthorized access in all 5 cross-session tests. Exact-action checks and transactional idempotency prevented duplicates in 10 repeated approval attempts. These are bounded test results for simulated actions.

Hosted on HF CPU Basic, it connects to Neon via certificate-verified WSS on port 443. Checkpoints survive container restarts; sessions still expire. The dated deployment report records successful live workflows, duplicate-approval protection and restart checks. It is a small fictional demo, not production load qualification; free services may sleep.

Open SupportDesk demo ↗

DocVerify

Extract uncertain fields. Check the numbers. Review the differences.

  • OCR
  • Pydantic
  • Decimal
  • FastAPI
Try DocVerify ↗
Architecture, evidence & limitations

DocVerify is my AI-assisted independent invoice/PO review application: FastAPI, React/TypeScript, PDF text/rendering, Tesseract OCR, Pydantic model extraction, Decimal arithmetic, SQLite, LiteLLM and metadata-only Langfuse. AI extracts uncertain fields; backend code checks totals and conservative line matches. Review preserves originals, revisions, source pages and JSON/CSV exports.

The public HF CPU demo offers 20 fictional pairs, not arbitrary uploads. Local deterministic and labelled-fixture checks passed. OCR/model errors and temporary review storage remain limitations; no currency conversion or tax interpretation.

Open DocVerify demo ↗

Skills with evidence

The tools, and where I use them.

Professional experience, independent implementation
and learning are different kinds of evidence.

Experience

Oracle · Application Software Engineer 2

At Oracle since August 2023 · Primavera P6 Professional · HyderabadRead the context

I’m an Application Software Engineer 2 at Oracle in Hyderabad, working on Primavera P6 Professional since August 2023. I have 3+ years of experience building and deploying generative AI applications used by others. I delivered 6 features and resolved 150+ defects during full-time employment, separate from my internship contributions.

I strengthened import-type and user-privilege authorization, enhanced imports with project information such as Responsible Manager, and contributed to Delphi-to-C# modernization. I raised recurring average build success from about 50% to 90% by addressing build-machine failures; I also fixed Build Verification Test failures and authored FitNesse regression cases.

I work directly with customers to gather requirements, explain technical solutions, demonstrate features, and discuss troubleshooting and workarounds on calls.

Oracle · Project Intern

January–July 2023 · Remote · 50+ defects resolved · 2 components migrated · 1 feature contributionRead the context

I worked remotely as an Oracle Project Intern from January to July 2023 on Primavera P6 Professional. I resolved 50+ defects in its C# and Delphi codebase, including critical bugs, migrated 2 existing components from Delphi to C#, and contributed to developing 1 feature. These contributions are separate from my full-time totals.

C-DAC · Summer Intern

June–July 2022 · Top-10 NIT Silchar selection · Team-built MERN mortality analytics for MaharashtraRead the context

I was selected among the top 10 NIT Silchar students for a C-DAC internship spanning June-July 2022. I co-developed a MERN-based mortality analytics system for the Maharashtra government to support pandemic outbreak monitoring and vaccination-effectiveness analysis. The system used MongoDB, Express, React and Node.js to digitize death registration and support analysis. This was team development with equal contributions across all parts; the six-week internship spanned those two months.

Education & learning

Still building on that foundation.

NIT Silchar

B.Tech, Electronics & Instrumentation Engineering · 2019-2023 · CGPA 9.14Read the context

I completed a B.Tech in Electronics and Instrumentation Engineering from NIT Silchar (July 2019-May 2023), with a 9.14 CGPA. I studied Class 12 (Science) at Pragya Academy, Jorhat (2017-2019), under the SEBA Board, Assam, and passed with 85.6%. I completed my schooling at Don Bosco High School, Lichubari, Jorhat, in 2017 with 83.83%, under the SEBA Board.

System Design Simplified - High Level Design

Gaurav Sen · InterviewReady · CompletedRead the context

I completed System Design Simplified - High Level Design by Gaurav Sen on InterviewReady, confirmed on 23 September 2026. This supports my system design/HLD learning and has helped me develop my projects and contribute to my work at Oracle.

AI, programming and full-stack learning

Udemy learning in AI engineering, programming, machine learning and full-stack development.Read the context

I took these Ed Donner courses on Udemy: AI Builder: Create Agents, Voice Agents, & Automations in n8n; AI Coder: Complete Claude Code & Coding Agents Course; AI Engineer Core Track: LLM Engineering, RAG, QLoRA, Agents; and AI Engineer Agentic Track: The Complete Agent & MCP Course. These are learning records, separate from production work.

Other Udemy courses I took: 100 Days of Code: The Complete Python Pro Bootcamp (Angela Yu); Mastering Data Structures & Algorithms using C and C++ (Abdul Bari); Machine Learning A-Z [2026]: ML, DL, AI with AWS, Python & R and Deep Learning A-Z [2026]: DL, AI in Python & AWS + LLM Prize (Kirill Eremenko and Hadelin de Ponteves); and The Complete Full-Stack Web Development Bootcamp (Angela Yu).

Institute work

Institute projects and achievements.

Projects and achievements

Projects, competitions and college contributionsRead the context

SDocs (June-July 2022) was a React document-management team project built during my undergraduate studies, with equal contributions across all parts from everyone on my team. It was selected among 15+ teams. Keeper (January-February 2022) used React/local storage for notes; Classroom (December 2021-January 2022) used Node, Express, EJS and MongoDB; Secrets (January 2022) supported anonymous sharing and authentication, with hashed passwords and 100+ historical registrations. All three were historically deployed.

In 2022, I ranked in the top 10 on GeeksforGeeks at my institute, placed 1,231 in LeetCode Biweekly Contest 83 and 131 in the July 2022 GFG Amazon Alexa Hiring Challenge (out of 22k+ candidates), and was selected for Amazon ML Summer School 2022. My college roles included Coding Club technical member, RoboBuild event organizer, Incandescence publicity/content team member and IGNITS publication team member. Our team also won the RoboBuild competition in February 2020.

Programming work

My programming background.

Programming background

Languages, tools and computer science foundationsRead the context

My professional product work includes C# and Delphi for debugging, features, performance improvement and modernization, and FitNesse for test cases and regression authoring. I prefer C++ for data structures and algorithms. Earlier web work includes HTML, CSS, JavaScript, React, Node.js, Express, EJS, MongoDB and MySQL. Coursework includes DSA, OOP, DBMS, operating systems and computer networks. My skills also include C, SQL and Git. In personal projects outside Oracle, I have worked with AWS S3, SageMaker and Bedrock and created Docker containers. My BugTrace retrieval uses OCI.

Continue the conversation

Interested in reliable software and practical AI? Connect with me directly.

LinkedIn ↗gaurabdas043@gmail.com

Recruitment

A Note to Recruiters

Recruiter details

AI software engineering and Cloud AI Engineer opportunities · Targeting SDE2 or SDE3 · One-month noticeRead the context

I am based in Hyderabad and interested in AI software engineering and Cloud AI Engineer opportunities, targeting SDE2 or SDE3 roles, combining enterprise application development with practical AI systems. My current Oracle title is Application Software Engineer 2; AI Software Engineer is my professional positioning. SDE2 and SDE3 are my target levels, and employer leveling conventions vary. Notice period is one month only. I’m open to opportunities in India and internationally, including remote, onsite and hybrid roles. Employer visa sponsorship is preferred.

Reviewed introductions you can read without using AI.

Quick introduction

I’m an Application Software Engineer 2 at Oracle, working on Primavera P6 Professional. I have 3+ years of experience building and deploying generative AI applications used by others.

My work combines AI tools, enterprise software development, and direct collaboration with customers. I’m looking for SDE2 or SDE3 opportunities in AI software engineering, including Cloud AI Engineer roles.

AI tools at Oracle

I built and deployed BugTrace, a VM-hosted RAG tool that retrieves evidence from 5,300+ historical bugs for developer investigations. I also created a Codex marketplace with four plugins for BugDB authentication recovery, MCP-based TFS integration, IPS release automation and FitNesse test authoring. Both tools serve the same group of 7 developers and 5 QA engineers.

I collaborate with my team on the P6 AI chatbot, now in internal beta. I’m developing local HTML reports from LLM-generated instructions to keep report rows out of model requests and reduce token usage and cost. The reporting work is still in development.

Engineering achievement

My shipped performance improvement reduced first project opening from approximately 52 to 16 seconds and closing from 29-35 to 2-3 seconds in a customer production environment with a large dataset. The customer expressed satisfaction with the improvement. These measurements apply to that environment. I achieved these gains through query batching, deduplication, event-invalidated caching and deferred loading.

I delivered 6 features and resolved 150+ defects during full-time employment. My work includes import authorization, import enhancements and Delphi-to-C# modernization. By addressing build-machine failures, I raised recurring average build success from about 50% to 90%. I also work directly with customers on requirements, demonstrations and troubleshooting.

Explore a project

Explore PriceLab for QLoRA training and measured price-estimation improvement; IssueLens for hybrid RAG and MCP; SupportDesk for agents, approvals and PostgreSQL; or DocVerify for OCR and deterministic invoice checks. These are my AI-assisted independent projects. Public demos may sleep or reach free-provider limits.

Roles & locations

I’m interested in AI software engineering and Cloud AI Engineer opportunities, targeting SDE2 or SDE3 roles. I’m based in Hyderabad with a one-month notice period only, open to opportunities in India and internationally, including remote, onsite and hybrid roles. Employer visa sponsorship is preferred. You can reach me on LinkedIn or at gaurabdas043@gmail.com.

Technical skills

My professional product work uses C# and Delphi for features, debugging, performance improvement and modernization, and FitNesse for regression tests. I prefer C++ for data structures and algorithms.

My independent AI projects use Python, FastAPI, React, PostgreSQL, LangChain, LangGraph, OpenAI Agents SDK, LiteLLM, MCP and model fine-tuning. Earlier web work includes Node.js, Express, EJS, MongoDB and MySQL; my coursework includes DSA, OOP, DBMS, operating systems and computer networks.

My skills also include C, SQL and Git. I have used AWS S3, SageMaker and Bedrock in personal projects outside Oracle and created Docker containers. BugTrace uses OCI for its embedding service.

Engineering decisions

I separate model suggestions from authoritative checks: SupportDesk requires exact-action approval and transactional writes; DocVerify uses Decimal arithmetic after uncertain extraction; IssueLens combines keyword and vector evidence with citation validation. In SupportDesk, ownership checks blocked unauthorized access in all 5 cross-session tests, and exact-action checks with transactional idempotency prevented duplicates in 10 repeated approval attempts. These were simulated actions in bounded tests. These controls address specific risks rather than eliminating every model error.

For the P6 AI chatbot’s internal beta, I’m developing local HTML reporting from LLM-generated instructions. The goal is to keep report rows out of model requests and reduce token usage and cost.

Education & learning

I graduated from NIT Silchar in 2023 with a B.Tech in Electronics and Instrumentation Engineering and a 9.14 CGPA. I completed System Design Simplified - High Level Design by Gaurav Sen on InterviewReady, and that learning has helped me develop my projects and contribute to my work at Oracle.

I took Ed Donner's AI engineering courses on Udemy, covering agents, RAG, QLoRA and AI-assisted development, along with courses in Python, data structures and algorithms, machine learning, deep learning and full-stack development. These are learning records, separate from production work.

How it works

How this twin works

A portfolio you can talk to

A curated professional profile, instant topics and one bounded, read-only AI agent.Read the context

This AI-assisted portfolio uses Python, FastAPI, Pydantic, Jinja2 HTML, custom CSS and vanilla JavaScript. An OpenAI Agents SDK agent calls OpenRouter free routing, with one read-only approved-facts tool, at most two model calls and a 60-second overall timeout. Prepared answers work without inference. References must belong to facts supplied by the tool. No browsing, files or external actions are available to the model. Conversations are isolated and short-lived; quota counters persist in SQLite.

My five independent AI projects date from August 2023 onward. I directed and iterated on implementation with Codex assistance, personally executed guided Colab training and reviewed results. These are independent portfolio projects, not Oracle production systems. I can discuss documented architecture and tradeoffs.

It answers from a reviewed professional profile. It can explain my documented work, but cannot verify new claims, act on my behalf, or answer unrelated coding requests. References support review; they are not a guarantee of accuracy.

Privacy

A private-by-default visit

Your conversation and data

Conversation storage, model providers and usage limits.Read the context

This application keeps conversation history in server memory for up to one hour, without saving transcripts or using analytics trackers. Clear chat removes your session. Custom questions and recent conversation context are sent to OpenRouter and its model provider, whose own data policies apply. Request counts are stored to protect the free allowance.

Contact references

The twin cannot send messages or act on my behalf.

Oracle Cloud certifications

I earned the Oracle Cloud Infrastructure 2025 Certified AI Foundations Associate and Oracle Cloud Infrastructure 2025 Certified Foundations Associate certifications in March 2025.