Career Log
Replaying the build history
Each role, the systems shipped, the decisions made, and what they taught.
Senior AI/ML Engineer
EcoRatings
Own production LLM systems end-to-end — from FastAPI service layers through LangGraph agent orchestration to Dockerized AWS deployment, monitoring, and evaluation.
Shipped
- Architected and deployed Dockerized FastAPI AI services on AWS ECS Fargate with autoscaling, health checks, and CloudWatch monitoring, serving live RAG and agent workloads.
- Built citation-grounded RAG pipelines over domain documents using pgvector + Chroma with hybrid retrieval, metadata filtering, and reranking, backed by S3 and AWS ingestion workflows.
- Designed LangGraph multi-step agentic workflows with tool execution, state management, validation loops, and human-review checkpoints.
- Built LLM evaluation harnesses measuring correctness, retrieval quality, hallucination risk, and schema adherence to gate releases.
Decisions
- Fallback model routing across OpenAI, Anthropic, and open-weight models to balance cost, latency, and accuracy.
- Hardened services with API auth, throttling, retries, timeouts, and schema validation.
- Made eval harnesses a release gate to catch regressions before deploy.
Lessons
- Production AI reliability is mostly disciplined software engineering — routing, retries, and evals matter more than model choice.
- Human-review checkpoints are cheap insurance for high-stakes agent actions.
FastAPILangGraphpgvectorChromaAWS ECS FargateS3CloudWatchOpenAIAnthropic
Artificial Intelligence Engineer
Scale AI
Contract (hourly) AI engineer building LLM evaluation datasets and structured feedback loops — RLHF/SFT, LoRA/PEFT fine-tuning, and Python validation pipelines to raise baseline model quality.
Shipped
- Built LLM evaluation datasets for reasoning, safety, and instruction-following tasks.
- Designed structured RLHF and SFT feedback loops to improve baseline model performance.
- Fine-tuned transformer models using LoRA and PEFT, improving efficiency while reducing training cost.
- Developed Python validation pipelines for automated data quality and model output monitoring.
- Implemented context-aware evaluation workflows using Model Context Protocol (MCP).
- Collaborated with ML researchers and product teams to define alignment and evaluation metrics.
- Reduced noisy evaluation signals by 25% through improved model testing pipelines.
Decisions
- Standardized evaluation datasets around reasoning, safety, and instruction-following so alignment metrics stayed comparable across runs.
- Chose LoRA/PEFT over full fine-tuning to cut training cost while keeping quality gains.
- Automated data-quality and output monitoring in Python to catch noisy signals before they reached model scoring.
Lessons
- Clean, well-scoped evaluation data moves model quality more than clever training tricks.
- Parameter-efficient fine-tuning gets most of the benefit of full fine-tuning at a fraction of the cost.
PythonPyTorchLoRAPEFTRLHFSFTMCPLLM Evaluation
Software Engineer
JaiwebSoft Technologies
Built and deployed production backend services on GCP with FastAPI, Django, Docker, and Cloud Run — including async event-driven pipelines and hardened production configuration.
Shipped
- Deployed backend services on Cloud Run with autoscaling, health checks, and versioned revisions.
- Designed REST APIs with authentication, authorization, validation, pagination, and structured error handling.
- Modelled and optimized PostgreSQL/MySQL schemas on Cloud SQL — indexing, migrations, and query tuning.
- Built async processing with Cloud Functions, Pub/Sub, and scheduled jobs for ingestion, sync, and notifications.
Decisions
- Moved coupling-heavy work to event-driven Pub/Sub pipelines.
- Centralized secrets in Secret Manager with least-privilege service accounts.
- Added observability via Cloud Logging, Monitoring, and alerts.
Lessons
- Event-driven decoupling pays off the moment ingestion volume becomes unpredictable.
- Schema and index design decide whether a service scales gracefully or falls over.
FastAPIDjangoDockerCloud RunCloud SQLPub/SubPostgreSQLMySQL
B.Tech, CS (Artificial Intelligence)
IIIT Delhi
Computer Science with an Artificial Intelligence specialization — foundations in ML, systems, algorithms, and applied AI, alongside competitive programming.
Shipped
- Specialized coursework across machine learning, NLP, and systems.
- Top 4 finish in the Lumos BUIDL Hackathon among 30+ teams.
- Codeforces rating 1352 with 300+ DSA problems solved.
Decisions
- Leaned into applied AI and production systems over pure research.
- Built algorithmic depth through competitive programming.
Lessons
- Strong DSA fundamentals compound into better systems intuition.
- Shipping beats theorizing — hackathons taught me to scope and deliver fast.
PythonJavaC++SQLMachine LearningNLP