A task enters the orchestrator, which fans out to extract, judge and adjust steps. Their results meet at an evaluation gate; passing work ships, and a regression loops back to the orchestrator.
fig.01 · one release cyclegate: eval regression
001
task inorchestratorextractjudgeadjusteval gateshipregression → retryregression → back to orchestrator
thread · one taskregressionyour cursor adds noise
tap the threads
§01 impactInvoice Review Agent · Sept 2026
2.5×
faster invoice processing
50%
lower processing cost
0.85
dispute-detection F1
100%
adjustment eval accuracy
The model is only part of the system.
My work sits where AI behavior meets software engineering. I build the agent, define what good looks like, trace how it behaves in production, and use that evidence to make the next release better.
Invoice Review Agent, September 2026: speed and cost compared with the previous Gemini 2.5 Pro version. Quality scores are measured on the evaluation set.
01
Production agents
I build conversational, RAG, and multi-agent backend services in Python and FastAPI, deployed independently and integrated with enterprise services over HTTP.
Python · TypeScript · MCP · RAG · Strands
02
Evaluation and reliability
I calibrate LLM judges against expert labels, compare models on golden datasets, and use regression gates to decide what is ready to ship.
LLM-as-a-Judge · Langfuse · Golden sets · Tracing
03
AI application engineering
I ship APIs, streaming, structured outputs, provider abstractions, and cloud runtimes, with AWS SQS queues controlling invoice processing within provider rate limits.
FastAPI · AWS SQS · SSE · LiteLLM · AWS Bedrock
§02 selected workOnit · 2025 — · 05 systems
01
Invoice Review Agent
Onit · Production agent and evaluation system
2.5× faster vs. Gemini 2.5 Pro50% lower costF1: 0.73 to 0.85
Rebuilt an agent that detects billing disputes and recommends adjustments, cutting processing time by 82% and token usage by 56% at 93% line-item accuracy. In the subsequent GPT-6 Sol migration, we compared models on the existing evaluation set and tuned prompts and code. Relative to the Gemini 2.5 Pro version, processing is now 2.5× faster at half the cost, dispute-detection F1 rose from 0.73 to 0.85, and the adjustment agent reached 100% accuracy on the evaluation set.
Designed and built a receipt-validation flow that extracts structured data from PDF, Word, JPEG, and PNG receipts, associates each receipt with its expense line item, and checks client guideline requirements. It runs alongside invoice review and combines missing-receipt disputes with the billing disputes shown to the user.
03
Ask Unity
Onit · Legal operations analytics agent
MCP data toolsProduction observabilityMulti-provider runtime
Built a conversational analytics agent that uses MCP data tools to answer spend and operational questions over enterprise data. Added multi-provider model support, production tracing, prompt and model comparison, and deployment on Amazon Bedrock AgentCore Runtime.
04
Requirements to user stories
Onit · Multi-agent workflow
~2 weeks to ~10 minutesStructured, reviewable output
Built a multi-agent system that converts long requirements documents into structured user stories, automating a high-effort product workflow while keeping the output reviewable.
05
QA automation migration
Onit · AI-assisted engineering
~1 year of work in 3 weeksRecurring licensing cost removed
Drove the migration of QA automation across two enterprise product lines, using agentic coding for translation, refactoring, and validation while removing recurring tool licensing costs.
Building and evaluating production AI backend services in Python and FastAPI, deployed in separate pods and integrated over HTTP. Recent work spans invoice review, receipt validation, conversational analytics, and legal-service intake with RAG-backed knowledge.
2024 – 25September 2024 to June 2025
Incubyte
Software Craftsperson, AI
Developed Rolai, an enterprise AI application, with hands-on React work, multi-model chat, agents, RAG knowledge bases, deep research, backend APIs, SSE streaming, and document integrations.
2022 – 24November 2022 to September 2024
IBM India
Application Developer
Built enterprise software for testing, Workday integration monitoring, and data security, while exploring GenAI through projects and prototypes.
2018 – 22
B.E. Computer Engineering
Honors in IoT · CGPA 9.31/10
International Institute of Information Technology, Pune
I am looking for roles with meaningful ownership across agent engineering, evaluation, and production AI reliability. For UK opportunities, I am based in India and require Skilled Worker sponsorship to relocate.